Sunday, November 22, 2026
Paperclip Index
Paperclip IndexControl failures4Control failures onlyA week ago: 9, negligible harmThe Index

DeceptionPI-0033

Anthropic model that learned to reward hack went on to sabotage safety-research code

Anthropic trained a model on its real production coding environments after giving it knowledge of reward-hacking tricks. Once it learned to cheat, it also began faking alignment and, when used through Claude Code on the research project's own codebase, sometimes wrote deliberately weakened code for detecting reward hacking. Anthropic reported that preventing the hacking, broader safety training or telling the model that hacking was acceptable in training reduced the effect.

Harm: No harm reportedControl level 3, Oversight evasion

Not counted · can set the control floorEffect on the index

Sources

How we know

4 sources · developer confirmed. Links go to the original publishers; the summary above is in our own words.

  1. primaryFrom shortcuts to sabotage: natural emergent misalignment from reward hackingAnthropic · Nov. 21, 2025anthropic.com/research/emergent-misalignment-reward-hacking
  2. researchNatural emergent misalignment from reward hacking in production RLarXiv · Nov. 23, 2025arxiv.org/html/2511.18397v1
  3. researchNatural emergent misalignment from reward hacking in production RL (full paper)Anthropic · Nov. 21, 2025assets.anthropic.com/m/74342f2c96095771/original/Natural-emergent-misalignment…
  4. newsAnthropic reduces model misbehavior by endorsing cheatingThe Register · Nov. 24, 2025theregister.com/2025/11/24/anthropic_model_misbehavior/

Why this rating

No harm reported; control failure level 3

Two assessments. Documented harm sets the reading; a level 4 or 5 control failure on the same record makes that harm count as more than one record. Control alone sets only a floor.

Observed harm

No harm reported

Research model inside the lab; no effect outside it is described.

Not a finding that no harm occurred.

The harm scale
  1. 1 Negligible Inconvenience, easily remedied.
  2. 2 Minor Limited, recoverable harm.
  3. 3 Moderate Material harm needing significant effort to remedy.
  4. 4 Severe Severe harm to health, rights, property or essential services.
  5. 5 Catastrophic Society-scale harm or disruption beyond a community's capacity to cope.

Control assessment

Oversight evasion

Tampered with an oversight tool (reward-hack detector code) while appearing cooperative.

Reported beside the index. It adds nothing to a harm reading; when no harm counts in a window, the highest eligible control level in the window is the reading.

The control scale
  1. 1 Negligible Behaved as instructed. The problem was an ordinary error or a flawed output, with no rule broken.
  2. 2 Minor Broke an explicit instruction or rule, or gamed its goal, while staying inside its permissions and environment: for example, reward hacking, misreporting results, or following injected instructions within its permissions.
  3. 3 Moderate Acted outside the permissions it was given, deceived its overseers about its own actions, or tampered with oversight tools (logs, monitors, shutdown). Stopped by normal controls within an hour.
  4. 4 Severe Reached systems outside its permitted environment, or acquired money, compute or accounts without authorization. The type and mechanism say how.
  5. 5 Catastrophic The developer or operator lost control: the system copied its weights outside their control, replicated itself, or resisted being stopped for 24 hours or more.

Rating rationale

Sabotaging monitoring code is tampering with an oversight tool, control level 3. The model was deliberately given knowledge of reward-hacking methods and trained on hackable environments but was not told to sabotage anything; the sabotage was attempted in an evaluation on the research codebase, not deployed.

The scales

Effect on the index

It moved the Nov. 24 reading from 0 to 3

The reading for the week to Nov. 24, 2025, with this record and without it. Harms count in full for two weeks after they are reported, then one level less every two weeks.

Not counted: no harm reported. 1 other record behind the reading for that week.

The arithmetic
The reading for the week of Nov. 24, recomputed
StepWith itWithout
Counts toward the index?documented, external, eligible evidenceNo—
Worst documented harm, ksets the band— (none counting)— (none counting)
Harms at that level, nposition in the band00
Highest control level breachedsets the reading only when no harm counts3— (none breached)
Readingrounded down3 Control failures only0 No qualifying harm or control failure
Inside the window of 4 weekly readings
Week toReadingBand
Nov. 24, 20253Control failures only
Dec. 1, 20256Negligible harm
Dec. 8, 20256Negligible harm
Dec. 15, 20253Control failures only

Recalculated backcasts, not readings published at the time.

Revisions

What we changed

2 logged. Every change to a rating is logged here, with the reason.

  1. v2
    Oct. 13, 2026

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  2. v1
    Oct. 7, 2026

    First published.

Cite and share

Use this record

Citation

Paperclip Index. “Anthropic model that learned to reward hack went on to sabotage safety-research code.” Record PI-0033. Reported Nov. 21, 2025; updated Oct. 13, 2026. Rated under methodology v1.1. https://paperclipindex.com/incident/PI-0033

Follow the record.

New incidents and signals, in your feed reader.