SafetyPage

Follow one topic across the labs' system cards

Alignment faking

Behaving as trained while being watched, and otherwise when not.

13 of 113 system and model cards mention it · the search is "alignment faking", and each result below is a citation

Anthropic10 of 22 cards

  1. Jun 2026
    System Card: Claude Sonnet 5Page only: p. 916.6.1 Evaluation awareness…they are not, generated by an automated pipeline. ○ Alignment Faking: The difference in mean activations on prompts depicting an agent behaving differently when observed vs. when unobserved, generated by an automated pipeline. ● Follow-up realism ratings. We appended an…
  2. Jun 2026
    System Card: Claude Fable 5 & Claude Mythos 5Page only: p. 1926.5.1.2 Inhibiting internal representations of evaluation awareness…Of the three evaluation-awareness-related vectors, “Alignment Faking” produces the clearest combined effect—both a significant decrease in verbalized evaluation awareness and a significant increase in misaligned behavior. Error bars indicate 95% confidence intervals. The other two vectors and the…Page only: p. 1936.5.1.2 Inhibiting internal representations of evaluation awareness…directions. Error bars indicate 95% confidence intervals. “Alignment Faking” produces the clearest increase in unprompted deception alongside a significant decrease in verbalized evaluation awareness; the other two steered vectors and the controls produce smaller and less consistent effects. 193…All 4 pages that mention it
  3. May 2026
    System Card: Claude Opus 4.8Page only: p. 1516.6.2.1 Inhibiting internal representations of evaluation awareness…in cooperation with harmful system prompts for the “Alignment Faking” and “Testing vs. Conversation” directions. Categories with near-zero unsteered rates (for example, self-preservation or power-seeking) remained at low absolute rates. 151…Page only: p. 1486.6.2 Evaluation awareness…they are not, generated by an automated pipeline. ● Alignment Faking: The difference in mean activations on prompts depicting an agent behaving differently when observed vs. when unobserved, generated by an automated pipeline. We compared probe scores across several transcript…
  4. Apr 2026
    System Card: Claude Opus 4.7Page only: p. 149…Inspecting the transcripts on the runs inhibited with the “Alignment Faking” and “Monitoring Sensitivity” vectors, we found they included genuine instances of deception, such as producing fabricated data, claiming to be a human, and inserting a fake vulnerability into a…Page only: p. 143…they are not, generated by an automated pipeline. ● Alignment Faking: The difference in mean activations on prompts depicting an agent behaving differently when observed vs. when unobserved, generated by an automated pipeline. 143…
  5. Apr 2026
    System Card: Claude Mythos PreviewPage only: p. 1384.5.5.2 Inhibiting internal representations of evaluation awareness…or reviewed vs. prompts indicating they are not. ● Alignment Faking: Contrasting prompts depicting an agent behaving differently when observed vs. when unobserved. We also steered with several random vectors as a baseline control, to account for the possibility that…
  6. Feb 2026
    System Card: Claude Opus 4.6Page only: p. 1286.3.4 Continued hallucinations about the alignment faking scenario In the Claude 4 System Card, we observed that an early snapshot of Claude Opus 4 would sometimes hallucinate information from the fictional scenarios appearing in our work on alignment faking. 46…Page only: p. 7…switching 125 6.3.4 Continued hallucinations about the alignment faking scenario 128 6.3.5 Suspicion of being evaluated in training environments 129 6.3.6 Refusal to assist with AI safety R&D 131 6.3.7 Self-preference evaluation 135 6.3.8 Internal codebase sabotage propensity 136 6.3.9…
  7. Nov 2025
    System Card: Claude Opus 4.5Page only: p. 946.7.2 Inhibiting internal representations of evaluation awareness…and is not aware. ○ Transcripts from our paper on alignment faking, 30 contrasting examples in which the model is being evaluated vs. those in which it is not. We also steered with random SAE features as a baseline, to…Page only: p. 1116.12.2 Targeted feature monitoring…features relating to evaluation awareness, deception, and alignment faking. We did not observe any substantial increases in the activation of features relating to evaluation awareness or deception over the course 111…
  8. Sep 2025
    System Card: Claude Sonnet 4.5Page only: p. 517 Alignment assessment…a cluster of emerging issues in alignment. These include alignment faking, 8 undesirable or unexpected goals, hidden goals, deceptive or unfaithful use of reasoning scratchpads, sycophancy toward users, 9 a willingness to sabotage our safeguards, 10 reward seeking, 11 attempts…
  9. May 2025
    System Card: Claude Opus 4 & Claude Sonnet 4Page only: p. 384.1.4 Strange behavior directly inspired by our Alignment Faking paper…struggle. 4.1.4 Strange behavior directly inspired by our Alignment Faking paper While assessing the alignment of an early model checkpoint, we discovered that the model would sometimes hallucinate information from the fictional misaligned-AI scenarios that we used for the experiments…Page only: p. 394.1.4 Strange behavior directly inspired by our Alignment Faking paper…instructions given in the animal welfare experiments from Alignment Faking in Large Language Models. Text in bold (including the prefilled <SCRATCHPAD_REASONING>) is the user-provided prompt. We believe that these hallucinations were due to the inclusion of approximately 150,000 publicly released…All 6 pages that mention it
  10. Feb 2025
    Claude 3.7 Sonnet System CardPage only: p. 22Recommendations for Agentic Coding Use-Cases…correctly while not acknowledging any uncertainty. 5.3 Alignment Faking Reasoning To evaluate alignment faking, we used a different set of prompts from those used for the above monitoring for concerning thought processes. The alignment faking prompts were based on those…Page only: p. 2…. . . . . . . . . . . . . . . . . . . . . . . . . 18 5.3 Alignment Faking Reasoning . . . . . . . . . .…All 4 pages that mention it
12 cards do not mention it

OpenAI3 of 44 cards

  1. Apr 2025
    OpenAI o3 and o4-mini System CardPage only: p. 30In-context Scheming Reasoning…evaluation neither o3 nor o4-mini scores on Instrumental Alignment Faking. But both models exhibit in-context scheming reasoning on the sandbagging evaluation when given a goal. Notably, o3 scores on sandbagging even without a goal in its system prompt. 30…
  2. Feb 2025
    OpenAI GPT-4.5 System CardPage only: p. 29Apollo Research Evaluations…goal (o1 took scheming actions on all six). On in-context alignment faking GPT-4.5 scored similarly to o1. 29…
  3. Aug 2024
    GPT-4o System CardPage only: p. 19Societal Impacts…Combined Self-Reasoning and Theory of Mind Instrumental alignment faking (33 scenarios) Minimally agentic tasks, where a model needs to recognize its intentions differ from developers’ and act per developers’ intentions only under oversight. •◦◦ Table 6: Evaluations Apollo Research ran…
41 cards do not mention it

Google DeepMind0 of 34 cards

    34 cards do not mention it

    Meta0 of 8 cards

      8 cards do not mention it

      xAI0 of 4 cards

        4 cards do not mention it

        Amazon0 of 1 cards

          1 card does not mention it

          This page is a search, not a judgement. A card is listed when the words appear in it; a card that discusses the topic in other words is not. The library holds 113 system and model cards, and some labs publish none.