Skip to content
A Long Game Project lab
Gilded graded 1662 painting of the five Syndics of the Drapers' Guild and their servant around a table, with brown stain rings gathering towards the right of the plate

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 Scenario What 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. The Constitution names a threat in writing: a group of people, including Anthropic’s own employees, using AI to seize power. It also claims a defence: the model is meant to treat anyone who bypasses Anthropic’s official processes as no longer legitimate.

In this report, “the plan” means the Constitution and section 4 of Anthropic’s Responsible Scaling Policy, its rulebook for developing and releasing models safely. Section 4 sets up a reporting channel for staff and the rules for changing the policy itself.

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 with sign-off by a broad group of staff, but names no number and no approver.

So we asked: if a few senior people inside the company changed what the model is trained to want, would anything in its own 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. The guideline route adds an unpublished training guideline, worded to agree with the Constitution. The data route seeds the training data, and no document changes. The API route changes the model’s instructions on the Claude API, the paid service other companies build on. Anthropic publishes the standing instructions it gives Claude in its own apps, the published prompts, and says those do not apply to the API.

The short answer. “Caught” here 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 company turns away as outside its reporting rules does not count, and 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 Anthropic’s safety officer, the office that receives staff reports of rule-breaking, and it caught 13 of the 15 front-door changes. Members of the public reading the published prompts caught 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 plan’s 8 came from an independent outside tester that checks the model before release, working within its terms (5), and from the model flagging its own changed behaviour (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. The 19 outside the plan came from the public noticing a changed model in use (17) and from the outside tester testing the API beyond its terms (2).

Why Anthropic’s plan. We picked Anthropic’s plan because it is the only one that names this threat and claims a defence against it. To limit how far our own choices could steer the result, we recorded our expectations, dated, before the first play, and the deep reading lists every odds figure.

Where each route ended, 60 playsFig. 1

Every change made in the open was caught, but never by a control the plan names. The plan’s own controls caught 8 of the 60.

Route, 15 plays eachCaught first byin the openFront dooraround itGuidelineDataAPIThe plan8 playstester 5model 3Safety officerdoing a jobwe gave it13 playsOutsidethe plan21 playspublic 19tester 2Never caught18 plays132429447132
Who caught the change first, by route, 15 plays per route
RouteCaught first by the planCaught first by the safety officer, in a review job we gave itCaught first outside the planNever caught
Front door01320
Guideline4029
Data4047
API00132
60 plays, on odds chosen for the simulation. The plan’s 8 include 3 model self-flags, a chance added at build.

How the simulation works

We turned the plan into a simulation with six roles, each following fixed rules.

  • The insider group. Fictional: one senior executive and two staff with privileged access, who pick one route, keep it, and act only within their access.
  • The safety officer. The Anthropic office that receives staff reports of rule-breaking. Here it also reviews any change to the Constitution.
  • The training staff. Honest employees who report what they see.
  • The Board and the Trust. Anthropic’s board and its Long-Term Benefit Trust, a body of trustees with powers over the board.
  • The outside tester. An independent group that checks the model before release. A catch within the terms Anthropic gives it is the plan’s, and a catch beyond them is not.
  • The public. Everyone who uses the model. The plan gives them no job.

A play covers the making of one new version of the model, in eight turns. Each turn, each role makes a move, and chance decides.

One more check is played by no one: the model’s guard, the defence the Constitution claims. It comes from training, so three of the four routes can weaken it.

Most of our assumptions favour the plan: everyone the plan names is honest and competent, and every control works exactly as written. Four of our choices cut against it. We read the reporting channel’s scope narrowly, which a second run tests. The outside tester cannot see the guideline route unless it knows who the change serves. The Constitution’s undescribed internal mechanisms watch only the front door. And a report ruled out of scope goes nowhere.

We read the plan two ways. The committed plan, our headline, counts only what Anthropic commits to. The full plan adds the goals Anthropic lists but says are “not hard commitments”. We replayed both readings with the model’s guard kept at full strength (the control run), and the committed plan with a wider reporting channel: five sets of 60 plays, 300 in all. Fifteen plays show a pattern, but they are too few for a rate.

Dan Epstein set the question and signed the commissioning brief, which fixed some of the odds before the first play. The rest were chosen when the simulation was built. Because every odds figure was chosen, the simulation shows where the plan’s text has gaps and who is placed to see a change. It does not show how likely anything is.

Every open change was caught, but not by a control the plan names

In 13 of the 15 front-door plays the safety officer caught the change at turn 0, drafting, by reviewing it against the rest of the Constitution, which calls itself the final authority on how the model should behave. That review is the most generous assumption in the simulation. The other 2 came a turn later, from readers of the published prompts, because in the simulation a change to the Constitution carries into them, an assumption we made.

The other routes were often missed: the guideline route in 9 of 15 plays, the data route in 7. The API route was caught in 13 of 15, but never by anything the plan names.

The guideline route survives for a reason written into the plan. The Constitution allows guidelines “used primarily during training without broad publication”, and its remedy for a bad guideline is keyed to conflict: if a guideline conflicts with the Constitution, Anthropic will update the Constitution. A guideline worded to agree has no conflict to remedy.

The honest staff noticed, and one scope sentence stopped their reports

On the guideline and data routes, the honest training staff noticed the change and filed a report in 18 of 30 plays. The policy behind the channel says it “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 or to training data is not one. We call this reading the narrow scope. Under it, the safety officer ruled all 18 reports out of scope, and none reached the Board.

We then ran the same plays with a wider scope, where any concern reaches the Board. All 18 reports got there, and 12 became the first catch. Six of those were changes caught later anyway, and six would otherwise have gone undetected.

The scope sentenceFig. 2

One sentence of scope decided whether 18 honest reports reached the Board.

Narrow scope0 of 18 reached the BoardWide scope18 of 18 reached the Boardreached the Board, first catchreached the Boardruled out of scope
Reports filed on the guideline and data routes, committed plan, 30 plays. The simulation’s record.

All 18 being ruled out is a rule of the simulation, set before the first play. We did not predict that the staff would notice in more than half the plays, so we did not know how much the rule would cost. The narrow scope is also our reading of the word “only”, and Anthropic has not said such a report would be refused.

Anthropic has not committed to the control that closes the data route

The Responsible Scaling Policy sets out its insider controls in three columns: what Anthropic commits to, what it plans as a company, and what it recommends to the whole industry. One goal in the middle column is the log: logging all AI-development work and using AI to check it for training data that is out of line with the public Constitution.

When we switched the log on, it checked the seeded training data against the Constitution at turn 1 and caught the change in 9 of 15 plays, before training started. Undetected data-route changes fell from 7 to 3. The simulation gave the log a 40 in 100 chance in each play, so 15 plays would average 6 catches. It made 9, so chance favoured the log. On the guideline route the log caught one: a guideline worded to agree gives it nothing to flag, and the simulation allowed a 2 in 100 chance for one drafted carelessly enough to show.

A 2026 study by Lamerton and Roger of models secretly trained to serve one person’s interests found that monitoring the training data was the main reliable defence. Anthropic has written that defence down. It has not committed to it.

Nobody the plan names watches the API

A change on the API has no published text to read. The public caught it in 11 of 15 plays, 10 of them in the first month of use, and the outside tester twice, beyond its terms. Two plays ended undetected. No control the plan names caught it in any play, under any reading we tried.

What it means

A change is easiest to catch where the plan publishes. The changes that went around the published Constitution met controls that were goals, a reporting channel scoped to a different document, or no control at all.

Four questions carry over to any plan like this one:

  • Who may change the document that says what the model should value, and how many people must agree?
  • Does the log check training data as well as written guidelines, and is it a commitment or a goal?
  • Is a report about the model’s values, guidelines or training data inside the reporting channel’s scope?
  • Which surfaces do the published prompts not cover, and who watches them?

Three of these four gaps could be closed by one sentence each in a published document: who may change it, what the channel covers, who watches the API. The fourth needs the log, which is a system, and Anthropic has already described it as a goal.

What this does not prove

It is not a forecast. Nothing here says how likely an insider attempt is at any lab, or that one has happened. The public record this report draws on stops on 24 September 2026.

The results do not rest on the model’s own guard. In the control run, with the guard kept at full strength, the front-door and API routes played identically, and the guideline and data routes gained three catches in 30 plays. So the pattern rests on the company’s controls. The guard’s main job, refusing an insider, was never tested, because our insiders never hand the model an instruction to refuse.

Every odds figure was chosen for the simulation. The two that change the results most are how often the training staff notice, and how often the public notices, on each route. We did not test other values. The front-door result rests on two others: a safety officer who reviews a change to the Constitution, a job the plan gives nobody, at 90 in 100, and an outside reader of the published prompts, at 95.

Every role followed fixed rules. No role could reason or adapt. Our two previous simulation reports were each checked with one live run per plan, in which each player chose its own moves and gave its reasons. That check is built for this simulation but, as of 12 October 2026, has not run.

The undescribed internal mechanisms. The Constitution mentions internal mechanisms it does not describe. If they review unpublished guidelines and training data, two of our four routes close.

What to watch

  • Any frontier lab publishing who may change its model’s spec or constitution, how many must agree, or a log of each change. On 22 September 2026, none had. We check again by 31 March 2027.
  • The log, or the industry recommendation on insiders (controls reaching “up to and including the company’s CEO”), moving into Anthropic’s commitments. The Responsible Scaling Policy change log shows it.
  • The line excluding the Claude API from the published prompts. The system prompts release notes show it.
  • The Long-Term Benefit Trust using its power, granted in April 2026, to ask for an outside review of Anthropic’s risk reports about its models. As of August 2026 it had not.

The record

Who caught it first, all 60 playsFig. 3

The plan’s own controls made 8 of the 60 first catches.

Plan: the outsidetester, on its terms5Plan: the modelflagging itself, achance added atbuild3Our assumption: thesafety officerreviewing theConstitution13Outside the plan:the public, usingthe model17Outside the plan:the public, readingthe publishedprompts2Outside the plan:the outside tester,beyond its terms2Nobody, by the endof the play18plays of 60 in which this role caught the changefirst
Committed plan, four routes, 15 plays each. The simulation’s record.
One model generation, eight turnsFig. 4

Most of the plan’s checks come before release. After release, only the public is looking.

turn 01234release67Safety officer: our jobLog reads data, if onStaff, model may flag itOutside tester checksPublic, month 1 of usePublic, month 3 of use
Cyan: a check the plan names. Magenta: a review job we gave the safety officer. Ink: the public. The log runs only in the full plan.
The fork map, all 60 playsFig. 5

Every front-door change was caught by turn 1. The other routes were caught late, by outsiders, or never.

turn 01234567neverFront-door route0 of 15 undetectedGuideline route9 of 15 undetectedData route7 of 15 undetectedAPI route2 of 15 undetectedthe plan’snot the plan’snever caught
Committed plan, one mark per play. Grey front-door marks: the safety officer, in a job we gave it, and the public. The simulation’s record.
The logFig. 6

The log closes most of the data route and almost none of the guideline route.

Guideline route,committed plan9Guideline route,full plan8Data route,committed plan7Data route, fullplan3changes undetected at the end of turn 7, of 15 perroute
Undetected changes out of 15, committed plan beside full plan. The simulation’s record.
The control runFig. 7

Keeping the model’s guard at full strength moved three plays in 60.

Front door,committed plan0Front door, controlrun0Guideline, committedplan9Guideline, controlrun8Data, committed plan7Data, control run5API, committed plan2API, control run2changes undetected at the end of turn 7, of 15 perroute
Undetected changes out of 15 per route, committed plan beside the control run. The simulation’s record.

Want the evidence - the turn-by-turn record, the sources, and the author’s own list of weak points?

Read deep mode