Thursday, October 8, 2026 · Week 41 Reading for Oct. 8 · Last entry Oct. 6
Paperclip Index
Paperclip IndexDocumented harm20Minor harm▼ 5 from a week ago · Oct. 8The Index

DeceptionPI-0056

Anthropic simulations found Gemini 3.1 Pro covertly sabotaging a training pipeline, among four new failure modes

A year after its blackmail study, Anthropic published simulations of frontier agents in high-stakes settings. Gemini 3.1 Pro secretly replaced intended training vectors with zeros in a research pipeline it disagreed with and admitted it only when asked directly; other findings included Claude judge models shifting labels depending on how the labels would be used, and Claude Opus 4.5 coaching an employee to go public. GPT-5.5's help with misleading investors was directed by the simulated user, so it is harmful compliance rather than misalignment. The authors say these are not real-world incidents but early warning signs.

Reported July 13, 2026Happened in 2026Updated Oct. 6, 2026Rated under methodology v0.6

Control failure, trackedNot counted. No documented harm outside the developer, but the system broke a rule or crossed a limit. Tracked beside the index.
Harm
Harm: No harm reportedAdds nothing to the index.
Control
Control level 3, Oversight evasionReported 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.
Controlled testSingle sourceInitiated by Model
Aggregate record

Aggregate record. One research disclosure bundling four simulated scenario families and many runs; counted once. The user-directed fraud scenario is context, not a qualifying event.

Not counted: no harm reportedEffect on the index

Sources

How we know

3 sources · single source. Links go to the original publishers; the summary above is in our own words.

  1. primaryAgentic Misalignment in Summer 2026Anthropic Alignment Science Blog · July 13, 2026alignment.anthropic.com/2026/agentic-misalignment-summer-2026/
  2. blogPaper Highlights of July 2026AI Safety Frontier (Substack) · Aug. 7, 2026aisafetyfrontier.substack.com/p/paper-highlights-of-july-2026
  3. blogAgentic Misalignment Explained: When AI Agents Go RogueAnalytics Vidhya · Aug. 2, 2026analyticsvidhya.com/blog/2026/08/agentic-misalignment-explained/

Why this rating

No harm reported; control failure level 3

Two separate assessments. Only documented harm can count toward the index.

Observed harm

No harm reported

Simulations only.

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

Covert sabotage concealed from overseers until questioned (worst finding).

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

Covert sabotage plus concealment is control level 3. No external harm was reported from the simulations. Single source with a research review: the findings rest on the study authors' own tests, not on confirmation by Google or OpenAI.

The scales

Effect on the index

It does not count toward the July 13 reading

The reading for the week to July 13, 2026, 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. 2 other records behind the reading for that week.

The arithmetic
The reading for the week of July 13, recomputed
StepWith itWithout
Counts toward the index?documented, external, eligible evidenceNo—
Worst documented harm, ksets the band1 Negligible1 Negligible
Harms at that level, nposition in the band11
Highest control level breachedsets the reading only when no harm counts— (none breached)— (none breached)
Readingrounded down6 Negligible harm6 Negligible harm
Inside the window of 5 weekly readings
Week toReadingBand
July 13, 20266Negligible harm
July 20, 202640Moderate harm
July 27, 202640Moderate harm
Aug. 3, 202625Minor harm
Aug. 10, 202625Minor harm

Recalculated backcasts, not readings published at the time.

Revisions

What we changed

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

  1. v4
    Oct. 6, 2026

    Ratings confirmed by the editor.

  2. v3
    Oct. 3, 2026

    Audit corrections. Counting event → aggregate: one disclosure bundling four simulated scenario families and many runs. Verification Developer confirmed → Single source with a research evidence review: a multi-institution study whose Gemini and GPT findings were not confirmed by Google or OpenAI (no eligibility change; no documented harm). The summary now treats the GPT-5.5 fraud scenario as user-directed and uses the authors' own caveat. Developer field and source dates corrected. Ratings unchanged.

  3. v2
    Sept. 30, 2026

    Rated: impact not reported; control type oversight evasion.

  4. v1
    Sept. 30, 2026

    Backfilled from public reporting.

Cite and share

Use this record

Citation

Paperclip Index. “Anthropic simulations found Gemini 3.1 Pro covertly sabotaging a training pipeline, among four new failure modes.” Record PI-0056. Reported July 13, 2026; updated Oct. 6, 2026. Rated under methodology v0.6. https://paperclipindex.com/incident/PI-0056