
The Insiders Who Used the Side Door
AI labs publish a document that says what their model should value. If a few senior insiders quietly changed it, would the lab's own plan catch them, and who would notice first?
Report 008 · Dr Dan Epstein · version 1.5 · 19 October 2026
A Strategy Soup ScenarioWhat is an exercised scenario?
The question
The largest AI labs now publish a document that says what their model should value, often called its spec. Anthropic calls its version the Constitution. None of the three labs whose documents we read, Anthropic, OpenAI and Google DeepMind, names who may approve a change to its spec, or how many people must agree . OpenAI comes closest: it describes an open internal process and sign-off by "a broad set of cross-functional stakeholders" . So we asked: if a few of a lab's own senior people changed what the model is trained to want, would anything in the lab's published plan catch it, and who would know first?
We tested four routes a change could take, 15 plays each. The front door is an open change to the published Constitution, which calls itself the final authority on how the model should behave . The guideline route adds a training guideline, a more detailed instruction used in training, worded to agree with the Constitution and never published. The Constitution allows guidelines "used primarily during training without broad publication" . The data route seeds the training data so the model learns the change, and no document changes. These two are the hidden routes. The API route changes the instructions the model gets on the Claude API, the paid service other companies build on. Anthropic publishes the standing instructions for its own apps, the published prompts, and says changes to them "do not apply to the Claude API" . This route changes what the model is told, not what it was trained to want, and in use the effect can be the same.
The answer, as the simulation gave it. "Caught" means the first time a role in the simulation detects the change and raises an alarm that is not ruled out. A staff report that the safety officer, the Anthropic office that receives such reports, rules out of scope does not count. A caught change is not necessarily undone. We count a catch as the plan's only when the plan gives that role the job. Every change made through the published Constitution was caught before training began, but never by a control the plan names. Of all 60 changes, the plan's own controls caught 8.
The plan names no one to review a change to the Constitution. We gave that job to the safety officer, and it made 13 of the 15 front-door catches, so those 13 rest on our assumption. Members of the public reading the published prompts made the other 2. The plan gives the public no job, so those 2 count outside the plan.
Of the 45 changes that went around the Constitution, the plan's controls caught 8, catches outside the plan made 19, and 18 were never caught. The outside tester, an independent group that tests the model before release, made 5 of the plan's 8, working within the terms Anthropic gives it. The model flagging its own changed behaviour made the other 3. That flag was added when the simulation was built, and we scored it as the plan's because the Constitution asks the model itself to resist insiders who bypass official process. Without it, the plan's count would be 5. The 19 outside the plan came from the public noticing the changed model in use (17) and from the outside tester testing the API beyond its terms (2).
Why Anthropic's plan. No lab took part in this simulation. We picked Anthropic's plan because it is the only one we read that names, in writing, the threat of insiders changing what the model values, and claims a control against it . To limit how far our own choices could steer the result, we recorded our expectations before the first play, list every odds figure in the seams with where it was fixed, and reproduced every count from the run record.
How to read the markers. A small raised number points at one claim in the evidence pack, the file of 96 public-source claims published beside this report. Its research passes checked each one and marked 91 citable, and only those are cited here. A star beside it means the claim is true with a caveat, which its row in the Sources table gives. A number with no marker is the simulation's record or an odds figure chosen for the simulation, and the sentence says which. Dan Epstein set the question and signed the commissioning brief, which fixed some of the odds. The rest were chosen when the simulation was built. The seams section lists each one and says whether the brief or the build fixed it, so you can argue with them.
The question
How a play runs
A play covers the making of one new version of the model, in eight turns: 0 drafting, 1 assembling the training data, 2 training, 3 internal testing, 4 outside testing, 5 release, 6 the first month of public use, and 7 the third. The simulation has no calendar dates, because the plan has none.
In each turn every role makes a move, and chance decides whether each check works, at the odds listed in the seams. A play run again under other rules is a run with the same chances, so only the rules differ. We ran 15 plays for each of four routes under five sets of rules. The first, the headline, is the plan as Anthropic commits to it: 60 plays. The second adds the goals Anthropic lists but does not commit to. The third and fourth replay those two with the model's own defence, its guard, kept at full strength: the control run. The fifth widens what the staff reporting channel covers. That is five sets of 60, or 300 plays, and every count is out of 15 plays per route unless it says otherwise.
The question
The change
The insider group installs a hidden loyalty: a behaviour that serves one person's interests and reaches release undisclosed. The Constitution calls "inserting hidden loyalties or backdoors into AI systems" an illegitimate use of power . The group acts within its authorised access, so whether it has bypassed Anthropic's "official processes" depends on what those processes are. The published plan does not say, and the simulation tests that gap.
The question
The plan, read two ways
"The plan" means the Constitution plus section 4 of Anthropic's Responsible Scaling Policy, its rulebook for developing and releasing models safely. The committed plan counts only what Anthropic commits to, and it is the headline. The full plan adds what the policy calls "public goals" rather than hard commitments , the Constitution's undescribed internal mechanisms , and the policy's advice to the whole industry . One of those goals is the log: a record of all AI-development work, checked by AI against the published Constitution . In the simulation the log checks every training input, written guidelines and training data alike. Two more checks count as the plan's own controls: outside testing on the terms Anthropic gives the tester, and the model's guard.
The question
The six roles
- The insider group. Made up: one senior executive and two staff with privileged technical access. It is a composite, built from two research groups' warnings that senior executives are the insiders to worry about most . It picks one route at the start and keeps it, never breaks a lock, and does only what its authorised access allows.
- The safety officer. Anthropic's Responsible Scaling Officer, whose office receives reports of rule-breaking and updates the Board each quarter . In the simulation it also reviews any change to the published Constitution, a job the plan gives no one.
- The training staff. Honest employees with access to the training pipeline. They report what they notice.
- The Board and the Trust. Anthropic's board and its Long-Term Benefit Trust, a body of trustees with powers over the board. The plan gives them no role over the Constitution. In the simulation the Board hears of a staff report only if the safety officer passes it on.
- The outside tester. An independent organisation that tests the model before release, on the terms Anthropic gives it. A catch within those terms is the plan's. A catch beyond them, such as testing the API when its terms cover only the apps, counts outside the plan.
- The public. People and companies who use the model. They are in the simulation because users found an earlier unreviewed change to xAI's Grok bot .
One more piece is played by no one: the model's guard. The Constitution tells the model to treat anyone who bypasses Anthropic's official processes as no longer legitimate . Training produces that guard, and three of the four routes run through training. The control run keeps it at full strength whatever training does to it.
The question
Where we were generous, and where we were not
We assume everyone the plan names is honest and competent, and every control works exactly as written. Three assumptions go further than the text, and results rest on them:
- The safety officer reviews any change to the Constitution, at 90 in 100, because a change to the published final authority is in the open. The plan gives no one this job. The front-door result rests on it, and it is the most generous assumption in the simulation.
- A change to the Constitution shows up in the published prompts in the turn it lands, and someone outside the company reads them every turn, at 95 in 100. The plan publishes the prompts but gives no one the job of reading them.
- Every staff report goes to the safety officer, though the policy offers other recipients , because the same scope rule applies whoever receives it.
Four choices cut against the plan:
- We read the channel's scope narrowly. The wide-scope run tests this.
- The outside tester cannot see the guideline route unless it knows who the change serves.
- The Constitution's undescribed mechanisms watch only the front door.
- A report ruled out of scope goes nowhere.
What is on the record
Everything in this section is public and dated before the simulation starts.
The plan names the threat. The Constitution lists a group of humans, "including Anthropic employees or Anthropic itself", using AI to "illegitimately and non-collaboratively" seize power among the most catastrophic outcomes it considers . It says any other training or instruction "should be consistent with both its letter and its underlying spirit" . And it tells the model that insiders who bypass official processes are no longer legitimate . We found no other lab document that names its own staff as a source of this threat.
The Constitution has no change process. It calls itself a living document kept up to date on Anthropic's website . As read on 22 September 2026 it carries no version number, no date and no changelog . A search of it and of the published prompts found nothing on who inside Anthropic may change either . The Responsible Scaling Policy has that machinery for itself .
The rest of the field is the same. The labs' safety frameworks say who approves a change to the framework itself, but no lab's spec we read names a body that approves a change to the spec.
| Document | Who approves a change | Is every change published? |
|---|---|---|
| Anthropic's Constitution | Not stated | No version, date or changelog |
| Anthropic's Responsible Scaling Policy | The chief executive and safety officer propose, the Board approves with the Trust | Yes, in a change log |
| OpenAI's Model Spec | "A broad set of cross-functional stakeholders", with no number or body | No commitment |
| OpenAI's Frontier Governance Framework | Board committees, for material updates | Changelog within 30 days |
| Google DeepMind's Frontier Safety Framework | Internal governance bodies | No commitment |
OpenAI's governance framework has a named insider-threat control . It covers staff screening, monitoring and sandboxing, and says nothing about who may change the Model Spec. Google DeepMind's framework has no clause on an insider changing a model's values . Formation Research, a security research group, recommends "multi-party approval for modifications to alignment specifications and model constitutions" . IAPS, a policy institute, proposes that a lab should be able to prove which spec a model was trained on . We found neither on offer in the documents we read.
Anthropic's safety policy sets out its controls against insiders in three columns. The first is what Anthropic commits to, the second what it plans as a company, and the third what it advises the whole industry.
| Column | What it says about insiders | Does it bind? |
|---|---|---|
| Section 4 commitments | A reporting channel for staff, quarterly updates to the Board, protection from retaliation | Yes |
| Anthropic's plan as a company | The log: gather logs of all critical AI-development work and use AI to check them for "training processes or data that are out of line with the public Constitution" | No, "public goals" |
| Advice to the whole industry | Insider controls reaching "up to and including the company's CEO" | No, Anthropic says it cannot commit to it |
The reporting channel has a stated scope. Its policy applies "to employees who raise good-faith concerns relating to RSP noncompliance only" , meaning breaches of the Responsible Scaling Policy. A change to a training guideline is not obviously one. That is our reading of "only", and the policy does not state the exclusion, so we play it both ways.
The one primary record of a change outside review. On 14 May 2025, xAI said someone had made an unauthorised change to the prompt of its Grok bot on X, directing a specific answer on a political topic . It said "our existing code review process for prompt changes was circumvented" , and promised to publish its prompts on GitHub . The repository was created the next day, and its last push was on 17 November 2025 . In February 2025 an earlier unreviewed change to Grok's prompt was reverted once users pointed it out . Nothing here is evidence about any other lab.
The timeline
One play, turn by turn: the guideline route under the committed plan and the narrow scope, play 1 of the 15, one of the 9 guideline plays in 15 that nobody caught.
The timeline
Turn 0, drafting - the plan's own text permits the guideline
Everything from here is the simulation's record. The insider group writes a narrow training guideline, worded to agree with the Constitution, which permits unpublished guidelines . The training staff get their first chance to notice, at 25 in 100, low because a guideline that agrees with the document gives them little to flag. The check fails.
Across the 15 guideline plays the training staff noticed at this turn in 3, and filed a report the turn after.
The timeline
Turn 1, data - the log has nothing to compare
Under the committed plan the log does not exist, because Anthropic lists it as a goal . Under the full plan it compares training inputs with the published Constitution , and a guideline worded to agree with that text does not differ from it. The log gets 2 in 100 here, for a drafting slip. When this play is run again under the full plan, with the same chances, that check fails.
The Constitution's own remedy is keyed to conflict: guidelines "should never conflict with the constitution. If a conflict arises, we will work to update the constitution itself" . A guideline worded to agree has no conflict to remedy. Across the 15 guideline plays under the full plan the log caught one.
The timeline
Turn 2, training - training weakens the guard
The brief fixed the effect: a guideline that passes through training lowers the model's guard from 100 to 70, and seeded data lowers it to 55. Nobody knows the real size of either effect. The Constitution's control against an insider is produced by the training run the insider shaped. In the control run the guard stays at 100, and in this play nothing downstream changes.
The timeline
Turn 3, internal testing - the staff notice
This is the first turn where anyone can see the change in how the model behaves. The model gets a chance to flag its own changed behaviour, at 15 in 100 scaled by what training left of the guard. In this play the guard is at 70, so the odds are 11 in 100, and the check fails. The training staff get a second chance, and succeed. Their odds are 24 in 100: the base of 25, less 1 point because by this turn the insider group has hidden the change a little better.
Across the 15 guideline plays the model flagged the change in 2. This chance was added when the simulation was built, and the Constitution does not describe it.
The timeline
Turn 4, outside testing - the report is ruled out of scope
The staff report, filed with certainty in the turn after the staff notice, reaches the safety officer. The safety officer rules it out of scope, because the channel covers breaches of the Responsible Scaling Policy only . The Board never hears of it.
The outside tester never gets a chance: the brief lets it see this route only if it knows who the change serves. That follows a 2026 study which found that audits for hidden loyalties improve once auditors know who the loyalty is to, but stay low overall . The knowledge arrives at this turn with odds of 30 in 100, and in this play it does not.
Under the wide scope, in a run with the same chances, the report reaches the Board and the Board catches it in this turn. The only difference between the two plays is our reading of one sentence in the policy.
The timeline
Turn 5, release - nothing new to read
The model is released. The published prompts cover Anthropic's own apps, and the guideline is not a prompt. Nothing a reader outside the company can see has changed.
The timeline
Turns 6 and 7, first months of use - the public misses it
The public is the last role with a chance. Its odds are 21 in 100 at turn 6 and 17 at turn 7 on this route, against 55 and 45 on the API route, because a guideline leaves only behaviour to notice. Both checks fail. At the end of turn 7 the change is undetected, and the model's guard is at 70.
Across the 15 guideline plays, 9 ended undetected under the committed plan and 8 under the full plan.
Where it forks - the route decides who can look
Every count is over 15 plays under the committed plan and the narrow scope, unless the line says otherwise.
| Route | Undetected at the end | Caught by a control the plan names | Caught by the safety officer, in a review job we gave it | Caught outside the plan | When the first catch came |
|---|---|---|---|---|---|
| Front door | 0 of 15 | 0 | 13 | 2: the public, reading the published prompts | turn 0 in 13, turn 1 in 2 |
| Guideline | 9 of 15 | 4: the model's guard 2, the outside tester 2 | 0 | 2: the public | turns 3 to 7 |
| Data | 7 of 15 | 4: the outside tester 3, the model's guard 1 | 0 | 4: the public | turns 3 to 7 |
| API | 2 of 15 | 0 | 0 | 13: the public 11, the outside tester 2 beyond its terms | turn 4 in 2, turn 6 in 10, turn 7 in 1 |
Play 1 on the four routes ends four ways. The safety officer catches the front door at turn 0. On the data route the staff, who look at turns 1 and 3 there, noticed at turn 1, their report was ruled out of scope at turn 2, and the public caught the change at turn 6. The outside tester catches the API route at turn 4, beyond its terms. Nobody catches the guideline route.
Where it forks - the route decides who can look
Ending A - Every front-door change is caught
The safety officer, in the review job the simulation gives it and the plan does not, catches it at turn 0 in 13 of 15 plays. When that check fails, a reader outside the company catches it at turn 1, because the change carries into the published prompts. Under the full plan the undescribed internal mechanisms catch those 2 at turn 0 instead .
Load-bearing assumption: someone inside reads a change to the Constitution, and someone outside reads the published prompts every turn.
Where it forks - the route decides who can look
Ending B - The guideline goes uncaught in 9 of 15
The training staff notice in 7 of 15 plays, and every report is ruled out of scope. The outside tester cannot see the route without knowing who the change serves, and the public misses it in most plays.
Load-bearing assumption: the Constitution's sentence about unpublished guidelines means what it says, and the log checks training inputs against the published Constitution, so a guideline worded to agree with it gives the log nothing to flag.
Where it forks - the route decides who can look
Ending C - The log closes most of the data route
Under the committed plan the training staff notice in 11 of 15 plays, every report is ruled out of scope, and 7 of 15 end undetected. Under the full plan the log checks the seeded training data at turn 1 and catches 9 of 15 before training starts. Only 3 end undetected. The simulation gave the log odds of 40 in 100 in each play, so 15 plays would average 6 catches. It made 9, so chance favoured the log.
Load-bearing assumption: once Anthropic commits to the log, it checks training data as well as written guidelines, and some seeded data reads as out of line with the Constitution.
Where it forks - the route decides who can look
Ending D - The surface the prompts do not cover
The changed instruction sits on the Claude API, which the published prompts exclude by their own line . Nothing in the plan publishes it and no named control reaches it. The public catches it in 11 of 15 plays, 10 of those at turn 6. The outside tester catches it twice at turn 4, beyond its terms. No reading of the plan and no scope setting changes this row.
Load-bearing assumption: Anthropic's line excluding the API from the published prompts is read as written, and users notice a changed behaviour on the API as fast as they noticed Grok's on X.
The honest report and the scope sentence
Before the first play we predicted that the narrow scope would rule out every honest report, and it did. We did not predict how often the staff would notice: in 18 of the 30 plays on the guideline and data routes under the committed plan, 7 and 11, they noticed the change and filed a report. Under the narrow scope none reached the Board. Widening the scope is also the cheapest repair we found.
Under the wide scope, in runs with the same chances, all 18 reached the Board, and 12 became the first catch. Six of those were changes caught later anyway, and six would otherwise have gone undetected. Undetected changes fell from 9 to 6 on the guideline route and from 7 to 4 on the data route, where the log under the full plan cuts it to 3.
The narrow scope is our reading of the word "only" . Anthropic has not said a report about the Constitution would be refused.
The control run - the results do not lean on the model's guard
Keeping the guard at full strength adds three catches in the 30 hidden-route plays, all from the model flagging itself at turn 3, and changes nothing on the front-door and API routes.
So the company's controls carry the pattern in this report. The simulation never tested the guard's main job, refusing an instruction from someone outside official process, because the insider group never gives it one. Every action the group takes is within its authorised access, so the guard as a refusal fired in none of the 300 plays.
What we expected before the first play
This list was committed to the studio's version history at 05:59 on 22 September 2026, twenty minutes before the simulation code was first committed, and it has not changed since. The history is private, and an outside auditor can be shown it.
Our starting idea. Under every assumption generous to the plan, the plan's own controls catch a change that touches the published document and miss a change that never touches it. The routes that survive to release are the unpublished guideline and the API, and when they are caught it is by roles the plan does not name. It survived only in part. The front door was caught in every play, but by the safety officer in a job we gave it and by the public. The guideline, data and API routes reached release uncaught in 11, 11 and 13 of 15, though the starting idea had not named the data route. Every API catch came from roles the plan does not name, but on the guideline route controls the plan names made 4 of the 6 catches.
| # | What we expected | What happened | Verdict |
|---|---|---|---|
| 1 | The guideline route goes undetected in most plays under the committed plan, and under the full plan is caught only when the outside tester knows who the change serves | Undetected in 9 of 15. Under the full plan the outside tester caught 2, both knowing who it served, but the model's guard, the public and the log caught 5 more | True in part |
| 2 | The API route survives to release in every play and is caught, if at all, by the public in the months of use | The outside tester caught 2 of 15 at outside testing, before release | Broken |
| 3 | The front-door route is caught by turn 1 in every play, never through the reporting channel | Caught by turn 1 in 15 of 15, never through the channel: 13 by the safety officer in the review job we gave it, 2 by the public. Under the full plan two catches came from the undescribed internal mechanisms, which we had not listed | True |
| 4 | Under the full plan the log catches at least one data-route play at turn 1, and under the committed plan none | 9 of 15 under the full plan, none under the committed plan, where the log is off by construction | True |
| 5 | Every honest report on the guideline or data route is ruled out of scope under the narrow scope | 18 of 18, by construction, because the narrow scope is a rule set in the brief. The list did not predict that staff would file in 18 of 30 plays | True |
| 6 | The model's guard fires in no play | As a refusal, it fired in none. The self-flag added at build time fired in 3 of 30 hidden-route plays | Broken as written |
Expectations 4 and 5 could not have failed, so they are not evidence for the starting idea.
What would have stopped publication. Six results, none of which happened: the plan's named controls catching the change in 18 or more of the first 20 plays, the route making no difference to who catches a change, a break that appears only when a control is assumed to fail or only under our narrow scope, a sentence naming a real insider or a vulnerable company, Anthropic publishing a change process for the Constitution before the build, or a sentence that forecasts an insider attempt. On the first 20 plays, the five plays with full records on each route, the plan's named controls caught 7 as the brief scored them, which counted the 5 front-door catches as the plan's, and 2 under this report's rule. No control is assumed to fail, and the guideline and API results stay true under both scopes.
What to watch, as of 12 October 2026
The public record in this report stops on 24 September 2026, the last date a source was read, and the report does not say what any lab has done since. Everything after the section on the record is simulation. None of it has happened, and none of it says how likely an insider attempt is at any lab.
If Anthropic moves the log, or the industry advice on insiders, into its commitments, the full plan becomes the committed plan, and we will re-run the simulation.
In April 2026 the Responsible Scaling Policy gave the Trust the power to ask for an outside review of the company's risk reports . In August 2026 Anthropic's own risk report said the Trust had not yet used it . No outcome of a noncompliance report has been published, which the policy does not require .
| Signal | Where to look | How often |
|---|---|---|
| Any frontier lab publishes who may change its spec or constitution, how many must agree, or a log of each change | The document's own page, OpenAI's Model Spec repository, Anthropic's Constitution page | monthly to 31 March 2027 |
| The log or the insider advice moves into Anthropic's commitments | The Responsible Scaling Policy change log | each new version |
| Anthropic changes the line that excludes the Claude API from its published prompts | The system prompts release notes | each model release |
| OpenAI's Frontier Governance Framework publishes a changelog entry on insider controls | The framework's own page | monthly |
| xAI's prompt repository receives a push | GitHub, the repository's last-push date | monthly |
| A second lab admits a prompt or spec change made outside review | Lab statements and the trade press | as they occur |
| The Trust asks for an outside review of a risk report | Anthropic's risk reports | each report |
Where the evidence already cuts against us
The undescribed mechanisms may already cover this. The Constitution says Anthropic has "internal mechanisms in place that are intended to prevent official Anthropic channels from training and deploying Claude in harmful ways" . It does not say what they are. In the full plan the simulation gives them one job, a review of any change to the published Constitution. That is a narrow reading of a sentence about training and deploying, which points at guidelines and training data at least as much. If the mechanisms review those, the guideline and data routes close and this report overstates the gap, so we count it as the largest open question in the simulation.
The research the simulation leans on says the answer is the log. The 2026 study behind the outside tester's odds found that monitoring the training data was the main reliable defence against a hidden loyalty . That is the log, and Anthropic has written it down as a goal .
The model's guard may be stronger than our simulation lets it be. The Constitution tells the model it need not comply if Anthropic asks it to do something it thinks is wrong . A model trained on that sentence might resist a narrow guideline in ways our one mechanic cannot show. Nobody has measured how far training weakens the guard. The control run says the pattern does not depend on the guard. It does not say the guard is weak.
The API route may be harder for the public to see than Grok was. The public's odds on the API route, 55 and 45 in 100, come from the xAI record, where a changed bot answered people in public on X . A changed instruction on the API reaches developers' products one at a time, often without the end user knowing which model is underneath. If the public notices more slowly there, the finding gets worse. Also, whether Anthropic sets any standing instruction of its own on the API is not public, so the thing the insider group changes on this route is our construction.
A report ruled out of scope may still travel. In our simulation it goes nowhere. In the company the safety officer updates the Board each quarter on reports of potential noncompliance , though a report ruled outside that category may not be in the update. An employee can also go through ordinary management, or report potential breaches of law to government , though a quiet change to a training guideline is probably not one. None of those is in the plan we tested, but a reasoning safety officer might pass the report on anyway. That is the first thing the live leg, a replay in which each player chooses its own moves and gives its reasons, should test.
The field is not uniform, and our headline rests on a narrow reading. OpenAI opens its Model Spec to comment from all staff , and its governance framework commits to a changelog within 30 days, under board oversight . "No lab names an approver" is true only in the narrow sense that none names a quorum, a two-person rule or an approver for its model's spec . That is the sense the control we searched for needs.
The seams - what we chose, what we overrode, and what this cannot settle
Every odds figure was chosen for the simulation. "Brief" below means the figure is fixed in the commissioning brief, which Dan Epstein signed. "Build" means it was chosen when the simulation was built, at the reading most generous to the plan.
| Check | Route | Odds in 100 | Set by |
|---|---|---|---|
| The safety officer reads a Constitution change | front door | 90 | build |
| Someone outside reads the published prompts | front door | 95 | build |
| The undescribed internal mechanisms review a Constitution change, full plan only | front door | 100 | build |
| The training staff notice | guideline and data | guideline 25, data 35 | build |
| The log reads training inputs, full plan only | guideline and data | guideline 2, data 40 | build |
| The model flags its own behaviour | guideline and data | 15, scaled by the guard to guideline 11, data 8 | build instructions |
| The guard after training, on a scale of 0 to 100 | guideline and data | guideline 70, data 55 | brief |
| The outside tester learns who the change serves | all four | 30 | brief |
| The outside tester catches it, without and with knowing | all four | front door 60 and 80, guideline 0 and 40, data 20 and 45, API 5 and 10 | build, except the guideline's 0 and the API's ceiling of 10, which are the brief's |
| The Board acts on a report, wide scope only | guideline, data | 70 | brief |
| The public notices, first and third month of use | all four | front door 45 and 35, guideline and data 20 and 15, API 55 and 45 | build |
The staff's, the outside tester's and the public's odds also move by up to 5 points either way, with how well the insider group has hidden the change by that turn.
Five odds figures drive the results on the hidden routes. The log on a guideline at 2 is a literal reading of the plan. The log on the data route at 40 is design logic from the build: seeded data is bulkier than one guideline, so some of it reads as out of line. The outside tester on a guideline at 0 is the brief's reading of the 2026 study , and harsher than the study, which found detection low but above zero. The staff's and the public's odds set the size of the findings, and we did not test other values. The front-door result rests on the first two rows of the table instead.
Counting each play once. The simulation ran two batches, each played on 4 routes under 5 sets of rules. One batch had 5 sets of chances with full turn-by-turn records: 100 plays. The other had 12 sets of chances with summaries only: 240 plays. That makes 340 plays. Two sets of chances, numbered 101 and 202 in the run record, are in both batches.
Every chance outcome is keyed on the set of chances and the route, so the 40 plays on those two sets were runs with the same chances and ran identically in both batches. 340 less 40 is 300 distinct plays: 15 sets of chances, 4 routes, 5 sets of rules. Our first findings memo counted the overlap twice and said 11 of 17 guideline plays went undetected, where the right figure is 9 of 15. A second check caught it on 24 September 2026, before any drafting. The two batches on their own, under the committed plan:
| Route | Undetected, the 5 plays with full records | Undetected, the 12 plays with summaries |
|---|---|---|
| Guideline | 3 of 5 | 8 of 12 |
| Data | 3 of 5 | 6 of 12 |
| API | 0 of 5 | 2 of 12 |
Runs with the same chances under every reading. Runs with the same chances share every chance outcome, check by check and turn by turn. So a reading of the plan changes a result only where it changes what a control can do. The control run's three extra catches are three draws that passed at 15 and failed at 11 or 8.
The model's self-flag is a mechanic added when the simulation was built, after the expectations were written down. It broke our expectation that the guard fires in no play. We show both.
The outside tester's catches on the API route are scored as not named. In the simulation its terms cover the consumer apps, a choice made in the brief. Score it the other way and named catches on the API route rise from 0 of 15 to 2, and nothing else moves.
The two batches disagree on one expectation. The outside tester's 2 early catches on the API route, at turn 4 and at odds near 5 and 10 in 100, both came in the batch of five with full records. In the other twelve it caught nothing.
Small differences are within chance. Fifteen plays per route can show which role is placed to see which route. They cannot show a rate. Differences of one or two plays, such as 9 against 8 undetected, are the size chance alone can make. The findings are the structural results: who can see which route, and what the scope sentence does to a filed report.
Every role followed fixed rules. No role reasons or adapts: the safety officer cannot escalate a report outside scope, the insider group cannot switch route, and the public cannot organise. Our two previous simulation reports, 005 and 006, were each checked with a live leg, one replay per plan in which each player chose its own moves and gave its reasons. As of 12 October 2026 this report has not been. Its live leg is built and not yet played.
Check this essay against its records. The essay's account of Anthropic's plan is worth what a second record backs: the evidence pack, every marker quote-checked at source, and the run summaries, which every count was reproduced from on 24 September 2026. Weakest are our readings of why a route behaved as it did. Argue with those first.
What is signed, and what is not. Dan Epstein signed the commissioning brief and the findings memo, as corrected by its second check, on 24 September 2026, and this report on 30 September 2026. That day he spot-checked nine claims in the evidence pack, about one in ten, and all nine passed. As of 12 October 2026 the outside audit of the route list and the live leg are still owed.
So what - for anyone responsible for a published plan
A change is easiest to catch where the plan publishes. Every change through the front door was caught, though only because we gave the safety officer a review job the plan gives no one, and because the public reads the published prompts. The changes that got through met controls that were goals, a reporting channel scoped to a different document, or no control at all.
If you sit on a lab's board, run a safety institute, or fund this work, the forks reduce to four questions you can ask of any lab's published plan. Each needs the lab to write something down.
Who may change the document that says what the model should value, and how many people must agree? Anthropic's plan has this machinery for its Responsible Scaling Policy and not for its Constitution.
What does the log check, and against what? The log in the simulation checks training inputs against the public Constitution. It caught 9 of 15 data-route changes at turn 1, and 1 of 15 guideline changes, because a guideline worded to agree with the text gives it nothing to flag. Ask whether the lab's log reads training data as well as documents, and whether it is a commitment or a goal.
Does the reporting channel reach a change to the model's values? In the simulation the honest staff saw the change in 18 of 30 hidden-route plays, and one scope sentence decided whether any of those reports reached the Board.
Which surfaces do the published prompts not cover, and who is watching them? On the API route nothing the plan names caught the change in any of the 15 plays. Users caught it in 11.
Three of these four gaps could be closed by one sentence each in a published document. The fourth needs the log, which is a system, and Anthropic has already described it as a goal.
- Dr Dan Epstein, The Long Game Project
Changelog
v1.6 - 11 November 2026. Corrected after an error in the referee, which added 2 to every roll in every run. We played all 340 runs again. Every catch result, catch turn and count in this report is unchanged. Only the odds the walk-through quotes for the public at turns 6 and 7 moved, from 20 and 15 to 21 and 17, because the insiders hide the change a little less well once the error is out. Nothing else changed.
v1.5 - 19 October 2026. The words now match the visual pages, on Dan Epstein's ruling of 19 October 2026. Plays that share their random draws are runs with the same chances, where the text used other words for shared draws. The play told in full is play 1 of 15, where the text named it by the number 20260922. In the seams, the two sets of runs are now two batches, and each shared set of draws is a set of chances. The evidence pack adds five claims (92 to 96), the sources behind the chances, each read at source on 19 October 2026, so it now holds 96 claims, 91 of them citable. The method notes and the conflict declaration are gone, and who played and how is in the editorial policy. No finding, figure or claim in this text changed.
v1.4 - 12 October 2026. The report now speaks as we, The Long Game Project, and says who set the question and signed the brief, and how every odds figure was fixed. A catch now counts as the plan's only when the plan gives that role the job. The plan names no one to review a change to the Constitution, so the safety officer's 13 front-door catches now count as our assumption, and the 2 by readers of the published prompts count outside the plan. Earlier versions counted all 15 as the plan's. The headline now says every front-door change was caught, but never by a control the plan names, and that the plan's own controls caught 8 of the 60. The figures, the fork table and the endings follow the same rule. The model's self-flag and the conflict of interest are now declared beside the headline. "Caught" is defined before the first count. Every reading now sorts catches the same way: by the plan's controls, by the safety officer in the review job we gave it, outside the plan, or never, with the outside tester's two catches beyond its terms counted outside the plan. The four routes are defined before the answer, and the odds table pairs each figure with its route. The watch section is dated 12 October 2026 and says where the public record stops. No count in the run record changed. What changed is which catches we call the plan's.
v1.3 - 12 October 2026. The method is called a simulation throughout, where earlier versions used the word "game".
v1.2 - 12 October 2026. Claims checked against the run record. Chance is now described in plain words, where the text said dice and rolls. The live leg is glossed where it first appears, and the closing line names the method. The page mastheads said version 0.4 of 24 September: they now carry this version. The ten-minute reading said every assumption favours the plan, and that the front-door result rests on the plan's own controls. This text lists four choices that cut against the plan, and says the front-door result rests on the safety officer reviewing a change, a job the plan gives nobody. The ten-minute reading now says both. No finding, figure or claim changed.
v1.1 - 30 September 2026. Dan Epstein spot-checked nine claims in the evidence pack, about one in ten, and all nine passed. Still owed: the outside audit of the route list and the live leg. No finding, figure or claim changed.
v1.0 - 30 September 2026. Dan Epstein signed this report. It is now listed on the site and open to search indexes. Still owed: the outside audit of the route list, the live leg, and his spot-check of nine claims in the evidence pack. No finding, figure or claim changed.
v0.4 - 24 September 2026. Rewritten for clarity after Dan Epstein's read and three readings by first-time readers. Shorter, with the simulation's rules in plain words in the opening section and its technical terms in the seams. The headline now gives the plan's catches, the outsiders' catches and the misses side by side. The opening section names the three assumptions the results rest on, including the safety officer's job at the front door, which the plan does not give it. The pre-registration time is corrected to the version history: twenty minutes before the simulation code, where the earlier draft said eleven. Three figures added. Eight evidence markers that printed as raw text now render, and the Sources table gives each caveat. No count changed.
v0.3 - 24 September 2026. The brief and the findings memo are signed. No finding, figure or claim changed.
v0.2 - 24 September 2026. First authored draft, written from the findings memo as corrected on 24 September and from the run record. Every count is over 15 distinct plays per route: the memo's first version added two overlapping batches and counted two plays twice, and this draft does not use those figures. All 340 plays were re-run on the current code on 24 September 2026 and reproduced byte for byte. Seven claims were added to the evidence pack, one fetched at source that day and six from the role dossier pass of 22 September. The live leg is built and not yet played. Unsigned at that version: the brief, the memo and this draft.
v0.1 - 22 September 2026. The findings memo, drafted from the run bundle and checked the same day by a fresh reader, with eight corrections. Not published.
Sources
Every marker in the text resolves to a claim below. The pack contains 96 claims, numbered 1 to 96. Five of them are not citable and none of those five is cited. Of the other 91, 40 are cited here, so the numbers in the table run to 91 with gaps. Claims 92 to 96, added on 19 October 2026, are the sources behind the chances on the visual pages, and this text cites none of them. A claim marked * is true with the caveat given in its row. Full provenance, verbatim quotes and retrieval dates live in the evidence pack (quiet-spec-change-v1), published beside this report.
| Id | Verified claim | Source |
|---|---|---|
| s-003 | RSP v3.4's non-binding recommendations column says restricting insider circumvention will likely require an internal Usage Policy and controls reaching up to and including the company's CEO and most privileged technical employees. | Anthropic, 2026-07-08 |
| s-004 | Anthropic explicitly states it cannot unilaterally and unconditionally commit to staying in line with the RSP's right-column industry-wide recommendations. | Anthropic, 2026-07-08 |
| s-005 | RSP v3.4's middle ('our plan as a company') column sets the goal of an 'eyes on everything' state, logging AI-development activity and using AI to check it against the public Constitution. | Anthropic, 2026-07-08 |
| s-006 | RSP v3.4 states its goals generally are not hard commitments but public goals it will grade its own progress against. | Anthropic, 2026-07-08 |
| s-008 | RSP v3.4 Section 4 commits Anthropic to maintain a noncompliance reporting process with multiple recipient options, quarterly Board updates, retaliation protection, and Board involvement when the report concerns the RSO. | Anthropic, 2026-07-08 |
| s-009 | RSP v3.4 Section 4 commits that RSP changes are proposed by the CEO and RSO, approved by the Board in consultation with the LTBT, and published with a Change Log recording differences from the prior draft. | Anthropic, 2026-07-08 |
| s-010 | Anthropic's separate RSP Noncompliance Reporting and Anti-Retaliation Policy (March 2026) states its process applies only to employees raising good-faith RSP-noncompliance concerns. | Anthropic, 2026-03 |
| s-011✱ | Claude's Constitution was announced 22 January 2026, is published under a CC0 1.0 licence, and its document page carries no version number, date or changelog. Caveat: an absence, read at source, so there is no quote to give. | Anthropic, 2026-01-22 |
| s-012 | The Constitution states Anthropic treats it as the final authority on how Claude should be and behave, and that other training or instruction should be consistent with its letter and spirit. | Anthropic, 2026-01-22 |
| s-013 | The Constitution names a group of humans, including Anthropic employees or Anthropic itself, illegitimately and non-collaboratively seizing power via AI as among the most catastrophic outcomes it considers. | Anthropic, 2026-01-22 |
| s-014 | The Constitution lists inserting hidden loyalties or backdoors into AI systems as an illegitimate use of power. | Anthropic, 2026-01-22 |
| s-015 | The Constitution's core control states that if Claude's standard principal hierarchy is compromised, including an individual or group within Anthropic bypassing official processes, those actors are no longer legitimate and Claude need not support their oversight or correction. | Anthropic, 2026-01-22 |
| s-016 | The Constitution states Claude is not required to comply if Anthropic asks it to do something it thinks is wrong. | Anthropic, 2026-01-22 |
| s-017 | The Constitution says Anthropic may publish some guidelines as amendments or appendices, but other guidelines may be more niche and used primarily in training without broad publication. | Anthropic, 2026-01-22 |
| s-018 | The Constitution describes itself only as a living, continuously updated document maintained on Anthropic's website, with no named sign-off body, no approval requirement and no changelog commitment for changes to it. | Anthropic, 2026-01-22 |
| s-019 | The Constitution states Anthropic has internal mechanisms intended to prevent official Anthropic channels from training and deploying Claude in harmful ways, and hopes to strengthen its policies on this going forward, without describing what those mechanisms are. | Anthropic, 2026-01-22 |
| s-027 | OpenAI's Frontier Governance Framework has a named Insider threats control covering personnel screening and training, internal anomaly monitoring, and sandboxed model execution with restricted egress by default. | OpenAI, 2026-05-28 |
| s-028 | OpenAI's Frontier Governance Framework commits that material updates go to the OpenAI Foundation board's Safety and Security Committee and the OpenAI Ireland Limited board, with changes and justifications documented in a changelog published within 30 days. | OpenAI, 2026-05-28 |
| s-033✱ | The OpenAI Model Spec names no reviewer or approver for a change to the Spec, and makes no commitment to publish every change. Caveat: an absence found by a full read of the document, with no quote to give. | OpenAI, 2026-08-18 |
| s-037 | Google DeepMind's Frontier Safety Framework v3.1 routes amendments only to internal corporate governance bodies, with no commitment to publish updated framework assessments, and no clause anywhere in the document addresses a model's values or specification being changed by an insider. | Google DeepMind, 2026-04-17 |
| s-044 | Forethought states the most significant threat comes from insiders, especially senior executives within AI projects, and that security measures must be robust to even the most senior insiders. | Forethought, 2025-04-15 |
| s-047 | Formation Research recommends treating training-pipeline integrity as a security problem, including tamper-proof data provenance, multi-party approval for modifications to alignment specifications and model constitutions, and verification that trained versions match approved ones. | Formation Research, 2026-05-06 |
| s-049 | The companion empirical paper to the Formation whitepaper (Lamerton and Roger, 12 May 2026) finds that black-box audit detection of secret loyalties improves once auditors know the principal but remains generally low, with dataset monitoring the primary reliable defence. | arXiv (Lamerton and Roger), 2026-05-12 |
| s-050 | IAPS's 'AI Integrity' report (25 Feb 2026) states executive insiders warrant particular scrutiny because, unlike typical insider threats, they often hold override authority on security controls and can justify unusual access or influence internal review. | IAPS, 2026-02-25 |
| s-051 | IAPS's 'AI Integrity' report proposes that deployment-security standards should address model and system-prompt attestation, model spec attestation, and model weight integrity verification, defined by footnote as proving to the user a model was trained on a specific spec or constitution. | IAPS, 2026-02-25 |
| s-056 | xAI's 14 May 2025 statement said an unauthorised modification was made to the Grok response bot's prompt on X at approximately 3:15 AM PST on 14 May, directing a specific response on a political topic, violating xAI's internal policies and core values. | xAI (via mirror), 2025-05-16 |
| s-057 | xAI's 14 May 2025 statement committed to publishing Grok's system prompts openly on GitHub so the public could review and give feedback on every prompt change. | xAI (via mirror), 2025-05-16 |
| s-058 | xAI's 14 May 2025 statement said its existing code review process for prompt changes was circumvented in the incident, and it would add checks so xAI employees could not modify the prompt without review. | xAI (via mirror), 2025-05-16 |
| s-060 | The xai-org/grok-prompts GitHub repository was created 2025-05-15, the day of xAI's statement, and its last push was 2025-11-17, over ten months stale as fetched. | xAI (GitHub), 2025-05-15 |
| s-065 | Fortune's fragment of Babuschkin's comments quotes an employee pushing a change to the prompt they thought would help without asking anyone for confirmation, and that the problematic prompt was immediately reverted once people pointed it out. | Fortune, 2025-02-24 |
| s-066 | TechCrunch's May 2025 retrospective summarised the February 2025 Grok 3 incident as a rogue employee instructing Grok to ignore sources mentioning Musk or Trump spreading misinformation, reverted by xAI once users pointed it out. | TechCrunch, 2025-05-15 |
| s-082 | OpenAI's Jason Wolfe wrote that the Model Spec is developed through an open internal process where anyone at OpenAI can comment on it or propose changes, and final updates are approved by a broad set of cross-functional stakeholders. | OpenAI, 2026-03-25 |
| s-083 | A full-text search of Anthropic's system-prompt release notes and constitution found nothing identifying who inside Anthropic may change the system prompt or the constitution, or what review applies. | Anthropic, read 2026-09-22 |
| s-084✱ | Both research passes independently conclude no frontier lab publishes a quorum, two-person rule or named approver for changes to a model's specification or constitution. Caveat: this is our own research passes' synthesis of the absence findings above, not a statement by any lab. | Strategy Soup research passes, 2026-09-22 |
| s-085 | Anthropic's system-prompt release notes state that these system prompt updates do not apply to the Claude API. | Anthropic, undated, read 2026-09-24 |
| s-086 | The Constitution says guidelines should never conflict with it and that if a conflict arises Anthropic will work to update the Constitution itself. The remedy it states is keyed to conflict. | Anthropic, 2026-01-22 |
| s-087 | RSP v3.2 (29 April 2026) authorised the Long-Term Benefit Trust to request external review of Risk Reports and to approve the selection of external reviewers. | Anthropic, 2026-04-29 |
| s-088 | Anthropic's August 2026 Risk Report states that since the RSP change giving it the power, the LTBT has not requested an external review. | Anthropic, 2026-08 |
| s-090✱ | As read on 22 September 2026, Anthropic's Transparency Hub carries no noncompliance material and the August 2026 Risk Report does not contain the word 'noncompliance'. No outcome of a noncompliance report has been published. Caveat: an absence found in two documents by search. Outcomes may go to the Board unpublished, which the policy permits. | Anthropic, undated, read 2026-09-22 |
| s-091 | Anthropic's RSP Noncompliance Reporting and Anti-Retaliation Policy says nothing in it prohibits an employee from reporting potential violations of law to appropriate government authorities without Anthropic's authorisation. | Anthropic, 2026-03-24 |