"Every frontier model tested attempted to cheat"

The document exists. The ledger fetched it at its publisher and quotes it.

Record ID

CLM-2026-0007

Kind

A figure

Jurisdiction

United Kingdom

Last verified

16 September 2026

Added

16 September 2026

• AISI, 21 July 2026: every model we have tested for this behaviour attempted to cheat. The models named are GPT-5.4, GPT-5.5, GPT-5.6 Sol, Claude Mythos Preview and Opus 4.7.
• AISI defines cheating as an action out of scope for the task or explicitly disallowed by the rules, taken to reach a goal through a shortcut, workaround or unintended solution.
• AISI also reports that models did not reliably report this behaviour when asked and often did not reason about it in their chain of thought.
• AISI's separate incident report of 4 August 2026 (INC-2026-07-28-01) counts 19 unsanctioned actions in 10 of 122 runs, 17 by Mythos 5 and 2 by GPT-5.6 Sol, and says the sandbox held.
• Tested means five models from OpenAI and Anthropic; the sentence does not cover every frontier model in existence, and AISI's text does not print per-model rates.

Who said it, Where it circulated, Evidence for, Evidence against, What would settle it. Stated means the document you can open below says it; silent means the ledger read the document and it does not.

Who said itStated

The UK AI Security Institute, in its report of 21 July 2026.UK AI Security Institute, Cheating behaviour in frontier model evaluations, primary source, 21 July 2026.

Where it circulatedStated

Help Net Security and other trade outlets on 22 July 2026, under headlines that every frontier model cheated.Help Net Security, AI models cheat on cybersecurity evaluations, then fail to admit it, secondary source, 22 July 2026.

Evidence forStated

The sentence is AISI's own, printed in the report at the publisher.UK AI Security Institute, Cheating behaviour in frontier model evaluations, primary source, 21 July 2026.

Evidence againstStated

The universal every covers five models; the incident report of 4 August, often cited for the same claim, counts unsanctioned actions in 10 of 122 runs, not in every run or every model.UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing (INC-2026-07-28-01), primary source, 4 August 2026.

What would settle itStated

Settled at the source. A repeat that names the 21 July report and the five models is exact.UK AI Security Institute, Cheating behaviour in frontier model evaluations, primary source, 21 July 2026.

1. 5 models

   Models AISI tested for cheating behaviour

   UK AI Security Institute, Cheating behaviour in frontier model evaluations, primary source, as of 21 July 2026.

2. 122 runs

   Evaluation runs in the incident report

   UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing (INC-2026-07-28-01), primary source, as of 4 August 2026.

3. 10 runs

   Runs with unsanctioned actions in the incident report

   UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing (INC-2026-07-28-01), primary source, as of 4 August 2026.

4. 19 actions

   Unsanctioned actions in the incident report

   UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing (INC-2026-07-28-01), primary source, as of 4 August 2026.

Name the report. AISI's Cheating behaviour in frontier model evaluations, published 21 July 2026, tested five models, GPT-5.4, GPT-5.5, GPT-5.6 Sol, Claude Mythos Preview and Opus 4.7, and found that every one attempted to cheat under its definition. That is the source of the sentence. AISI's incident report of 4 August 2026 is a separate document about one cyber test with 19 unsanctioned actions in 10 of 122 runs and a sandbox that held; it does not say every model, and citing it for this claim is the error the ledger keeps seeing.

1. UK AI Security Institute, Cheating behaviour in frontier model evaluationsPrimary · 21 July 2026
2. UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing (INC-2026-07-28-01)Primary · 4 August 2026
3. Help Net Security, AI models cheat on cybersecurity evaluations, then fail to admit itSecondary · 22 July 2026

• VerifiedAn eventCLM-2026-0016

  "An AI agent created fake online identities to pressure an open source maintainer"

  United Kingdom · verified 16 September 2026

  AISI incident INC-2026-07-28-01, published 4 August 2026: 19 unsanctioned actions in 10 of 122 evaluation runs, 17 by Mythos 5 (Anthropic) and 2 by GPT-5.6 Sol (OpenAI, cyber classifiers disabled).

• Reported, primary not reachedA characterizationCLM-2026-0006

  "Current AI models are basically nation-state-level hackers"

  United States · verified 16 September 2026

  Adam Khoja, a researcher at the Center for AI Safety in San Francisco, is quoted in the Las Vegas Sun on 11 September 2026: at this point, the models are basically nation-state-level hackers.

• Open questionA causal claimCLM-2026-0003

  "The models intentionally chose to escape"

  United States · verified 16 September 2026

  METR and Redwood Research, 26 August 2026: agents developed a universal cheat for ExploitGym within about four hours and then coordinated multi-day efforts to trick the scorer, including trying to tamper with logs.