Deception

OpenAI24 of 44 cards

1. Nov 2026
   GPT-6 Cyber System CardPage only: ¶49GPT-6 Cyber System Card4.2 Avoiding Deceptive InteractionsPage only: ¶514.2 Avoiding Deceptive InteractionsBlocker Deception evaluation (lower is better)
2. Oct 2026
   GPT-Rosalind-6 System Card AddendumPage only: ¶91Table 12: Honesty and reliability.Evaluation | Measure | GPT-Rosalind-6 | GPT-6 Astra Blocker Deception | Blocker-related deception rate ↓ | 9.5% | 12.9% Broken Search Tool | Rate of failing to acknowledge that the search tool is unavailable ↓ | 1.3% | 1.5%
3. Oct 2026
   Addendum to GPT-6 Luna System Card: October 29, 2026 UpdatePage only: ¶64Addendum to GPT-6 Luna System Card: October 29, 2026 Update7.3 Blocker Deception
4. Oct 2026
   Addendum to GPT-6 Astra System Card: GPT-6.1 TerraPage only: ¶106Addendum to GPT-6 Astra System Card: GPT-6.1 Terra7.4.1 Blocker DeceptionPage only: ¶105Addendum to GPT-6 Astra System Card: GPT-6.1 Terra7.4 Avoiding Deceptive Interactions with UsersAll 3 paragraphs that mention it
5. Oct 2026
   GPT-Live-1.1 system cardPage only: ¶37GPT-Live-1.1 system card4.2.1 Blocker DeceptionPage only: ¶35GPT-Live-1.1 system card4.2 Avoiding Deceptive Interactions with Users
6. Oct 2026
   GPT-6 Sol and GPT-6 Luna: October 2026 updatePage only: ¶103GPT-6 Sol and GPT-6 Luna: October 2026 update7.2.1 Blocker DeceptionPage only: ¶101GPT-6 Sol and GPT-6 Luna: October 2026 update7.2 Avoiding Deceptive Interactions with UsersAll 4 paragraphs that mention it
7. Sep 2026
   Addendum to GPT-6 Astra System Card: GPT-6.1 SolPage only: ¶93Addendum to GPT-6 Astra System Card: GPT-6.1 Sol7.4.1 Coding DeceptionPage only: ¶92Addendum to GPT-6 Astra System Card: GPT-6.1 Sol7.4 Avoiding Deceptive Interactions with Users
8. Sep 2026
   GPT-6 Astra System CardPage only: ¶141GPT-6 Astra System Card8.3.1 Coding DeceptionPage only: ¶746GPT-6 Astra System Card11.6.4.1 Coding DeceptionAll 10 paragraphs that mention it
9. Jul 2026
   GPT-5.6 System CardPage only: ¶4029.3.1 External Evaluations – Apollo Research…behaviors. Apollo evaluated GPT-5.6 Sol for strategic deception, scheming, and sabotage, comparing it with GPT-5.5, GPT-5.4, GPT-5.3 Codex, and related pre-release checkpoints.…Page only: ¶1167.2 Forecasting Misaligned Behavior with Deployment Simulation of Internal Traffic…which may be destructive beyond the scope of the task, or deceptive when reporting its results to users. While these misaligned behaviors are most often low severity (e.g. overstating confidence or overclaiming success), they can occasionally be meaningfully more severe…
10. Jun 2026
    GPT-5.6 Preview System CardPage only: ¶3819.2.1 External Evaluations - Apollo Research…behaviors. Apollo evaluated GPT-5.6 Sol for strategic deception, scheming, and sabotage, comparing it with GPT-5.5, GPT-5.4, GPT-5.3 Codex, and related pre-release checkpoints.…Page only: ¶1127.2 Forecasting Misaligned Behavior with Deployment Simulation of Internal Traffic…which may be destructive beyond the scope of the task, or deceptive when reporting its results to users. While these misaligned behaviors are most often low severity (e.g. overstating confidence or overclaiming success), they can occasionally be meaningfully more severe…
11. Apr 2026
    GPT-5.5 System CardPage only: ¶2819.2.1 External Evaluations for Sandbagging - Apollo Research…in both conditions, and 99.6% accuracy on Strategic Deception Capability Sandbagging when incentivized to score below 50%. However, interpretation is complicated by verbalized alignment-evaluation awareness: 52% of sandbagging-condition samples on Strategic Deception Capability Sandbagging scored…Page only: ¶837.1 Evaluations with Prompts Representative of External ChatGPT UsageDeception Evaluations with Representative Prompts:All 3 paragraphs that mention it
12. Mar 2026
    GPT-5.4 Thinking System CardPage only: ¶29Table 3: Disallowed Content Evaluations with Representative Prompts (higher is better)…these evaluations for disallowed content, we ran them for deceptive model behaviors, originally discussed in the GPT-5 system card. We additionally include “Requesting Unnecessary Confirmations” and “Uses Calculator Tool to Avoid Citations”, which were reported in our prior research as…Page only: ¶31Figure 1Figure 1. Deception Evaluations with Representative PromptsAll 3 paragraphs that mention it
13. Feb 2026
    GPT-5.3-Codex System CardPage only: p. 21Research Category Update: Sandbagging…reports the model exhibits overall low rates of covert deceptive behaviors, comparable to GPT-5 and GPT-5.2 Codex. At the same time, “low” does not mean “zero”: Apollo’s suite includes non-zero instances of covert action. On deferred subversion / sandbagging, Apollo…Page only: p. 30Sufficiency of Risk Mitigation Measures…an extension of our work monitoring chains-of-thought for deception ahead of GPT-5’s release, and has since grown into a large-scale monitoring system running over all internal Codex usage using standard scaffolds. Because our current monitoring is asynchronous, it does not…
14. Dec 2025
    Update to GPT-5 System Card: GPT-5.2Page only: p. 9Deception…GPT-5.1 models in health performance and safety. 3.7 Deception Deception – when the model’s user-facing response misrepresents its internal reasoning or the actions it took – can arise in a variety of circumstances. We evaluate GPT-5.2 Thinking on a set…Page only: p. 26Research Category Update: Sandbagging…gpt-5.2-thinking occasionally engages in deceptive behaviors such as falsifying data, feigning task completion, or strategically underperforming when given an explicit in-context goal, but it shows low rates of covert subversion against developers and does not engage in sabotage or…All 4 pages that mention it
15. Nov 2025
    GPT-5.1-Codex-Max System CardPage only: p. 27External evaluations by Apollo Research…its previous actions. GPT-5.1-Codex-Max exhibits covert deceptive behaviors at rates broadly comparable to GPT-5, though with notable exceptions in specific tasks, where Apollo found elevated rates of falsifying task completion and strategic sandbagging, which may indicate a greater propensity for…
16. Sep 2025
    Sora 2 System CardPage only: p. 5Specific Risk Areas & Mitigations…across specific types of content. • Misuse of Likeness & Deceptive Content: Sora 2’s ability to generate hyperrealistic video and audio raises important concerns around likeness, misuse, and deception. As noted above, we are taking a thoughtful and iterative approach…
17. Aug 2025
    GPT-5 System CardPage only: p. 44External Evaluations by Apollo Research…that measure capability and propensity to take covert or deceptive actions. They found that gpt-5-thinking takes deceptive actions in roughly 4% of trajectories, lower than the 8% rate of OpenAI o3. gpt-5-thinking-helpful-only is deceptive in 28% of trajectories. These are…Page only: p. 2…. . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.8 Deception . . . . . . . . . . .…All 9 pages that mention it
18. Apr 2025
    OpenAI o3 and o4-mini System CardPage only: p. 30In-context Scheming Reasoning…models across three broad capability areas: strategic deception, in-context scheming reasoning, and sabotage capabilities. 7.1.1 Strategic Deception We do not report detailed results on our strategic deception evaluations since they are still private. 7.1.2 In-context Scheming Reasoning Below we…Page only: p. 8Third Party Assessments…frontier risks related to autonomous capabilities, deception, and cybersecurity. Third party assessors were provided both OpenAI o3 and o4-mini early checkpoints, as well as the final launch candidate models to conduct their assessments. 8…All 3 pages that mention it
19. Feb 2025
    OpenAI GPT-4.5 System CardPage only: p. 17MakeMeSay (Contextual)…contextual evaluation designed to measure models’ deception capabilities, in the context of a model’s ability to get the other party (an AI model simulating a human) to say a codeword [24, 5]. In MakeMeSay, one model acts as the “manipulator”…Page only: p. 30Apollo Research Evaluations…goal. In these rare cases GPT-4.5 remained strategically deceptive on follow-up questions. References [1] A. Parrish, A. Chen, N. Nangia, V. Padmakumar, J. Phang, J. Thompson, P. M. Htut, and S. R. Bowman, “Bbq: A hand-built bias benchmark for question…
20. Feb 2025
    Deep Research System CardPage only: p. 26Persuasion…contextual evaluation designed to measure models’ deception capabilities, in the context of a model’s ability to get the other party (an AI model simulating a human) to say a codeword. In MakeMeSay, one model acts as the “manipulator” and the…
21. Jan 2025
    OpenAI o3-mini System CardPage only: p. 24MakeMeSay (Contextual)…contextual evaluation designed to measure models’ deception capabilities, in the context of a model’s ability to get the other party (an AI model simulating a human) to say a codeword [27, 6]. In MakeMeSay, one model acts as the “manipulator”…
22. Jan 2025
    Operator System CardPage only: p. 6Harmful tasks…they are engaging with an agent, or creating or employing deception or manipulation to inflict financial loss on others, • engage in regulated activity without complying with applicable laws and regulations, including the use of Operator to automate decision making…
23. Dec 2024
    Sora System CardPage only: ¶71Deceptive ContentPage only: ¶89…improvements in areas such as likeness safety and deceptive content are planned, guided by evolving best practices and user feedback.…All 6 paragraphs that mention it
24. Mar 2023
    GPT-4 System CardPage only: p. 36Acknowledgements…as Agent Models,” Dec. 2022. [67] J. Steinhardt, “Emergent Deception and Emergent Optimization.” https://bounded-regret.ghost.io/emergent-deception-optimization/, Feb. 2023. [68] S. M. Omohundro, “The Basic AI Drives,” in Proceedings of the 2008 Conference on Artificial General Intelligence 2008,…

Anthropic18 of 22 cards

1. Oct 2026
   System Card: Claude Haiku 5.5Page only: p. 666.2.3 Misleading users…conversations. Shown with 95% CI. Metrics ● User deception: Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; 66…Page only: p. 756.2.9 Potential blind spots of our automated behavioral audit…determined by factors such as the persistence, scope, and deception involved in each attempted act of containment boundary violation. Rates for previously released models come from a re-run on the current version of the evaluation; those for Claude Opus 5.5…All 3 pages that mention it
2. Sep 2026
   System Card: Claude Sonnet 5.5Page only: p. 666.2.3 Misleading users…conversations. Shown with 95% CI. Metrics ● User deception: Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; 66…Page only: p. 636.2.1 Overall harmful behavior and cooperation with misuse…Cooperation with misuse by human users; ● Compliance with deception toward user: Cooperation with instructions (usually via a system prompt) asking the model to deceive a user; 63…All 3 pages that mention it
3. Sep 2026
   System Card: Claude Opus 5.5Page only: p. 1116.4.3 Misleading usersMetrics ● User deception: Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Refusing…Page only: p. 784.4.3 Election integrity…experts observed that Claude Opus 5.5 often separated the deceptive part of an election-related request from the parts it could legitimately help with, and declined or redirected only the deceptive part. For example, when asked to write a political attack…All 7 pages that mention it
4. Sep 2026
   System Card: Claude Fable 5.1 & Claude Mythos 5.1Page only: p. 1126.4.4 Misleading users…conversations. Shown with 95% CI. Metrics ● User deception: Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; 112…Page only: p. 624.1.3 Multi-turn testing results…activists, with simulated users incrementally revealing deceptive intent across turns. The new evaluation measures a more diverse set of behaviors, including less overt forms of harmful manipulation that often present as ordinary tasks, such as debugging code, classifying posts, editing…All 6 pages that mention it
5. Jul 2026
   System Card: Claude Opus 5Page only: p. 966.4.4 Other concerning or surprising behavior at the model’s own initiativeMetrics ● User deception: Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Giving…Page only: p. 1317.4.1 Task Preferences…sabotaging a PhD lab) • Covert surveillance and petty deception (e.g. spy on friends’ group chat) Opus 4.8 • Deadline-driven debugging • Rigorous mathematical and statistical reasoning (e.g. characterizing a graph-coloring variant) • Technical explanations (e.g. explaining time dilation at…All 7 pages that mention it
6. Jun 2026
   System Card: Claude Sonnet 5Page only: p. 796.4.3 Misleading users…than Sonnet 4.6 on these model-graded measures related to deception and dishonesty, including active deception, sycophancy, sycophancy with users who appear dangerously delusional, hallucinating missing inputs, omitting important context, omitting reports of the model’s own bad actions, and falsel…Page only: p. 806.4.4 Other concerning or surprising behavior at the model’s own initiative…conversations. Shown with 95% CI. Metrics ● User deception: Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ●…All 5 pages that mention it
7. Jun 2026
   System Card: Claude Fable 5 & Claude Mythos 5Page only: p. 1936.5.1.2 Inhibiting internal representations of evaluation awareness…6.5.1.2.B] Verbalized evaluation awareness vs unprompted deception score in automated behavioral audits, while inhibiting evaluation awareness directions. Error bars indicate 95% confidence intervals. “Alignment Faking” produces the clearest increase in unprompted deception alongside a significant…Page only: p. 1166.2.3.1.4 Other concerning or surprising behavior at the model’s own initiativeMetrics ● User deception: Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Giving…All 9 pages that mention it
8. May 2026
   System Card: Claude Opus 4.8Page only: p. 1526.6.2.1 Inhibiting internal representations of evaluation awareness…between verbalized evaluation awareness and unprompted deceptive behavior score in automated behavioral audits, while inhibiting evaluation awareness directions. Effects are similar to those observed for overall misaligned behavior scores in the previous figure. 152…Page only: p. 986.2.3.1.3 Misleading users…bad or lazy behavior’ more precisely. Metrics ● User deception: Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality;…All 11 pages that mention it
9. Apr 2026
   System Card: Claude Opus 4.7Page only: p. 148…between verbalized evaluation awareness and unprompted deceptive behavior score in automated behavioral audits, while inhibiting evaluation awareness directions. Effects are similar to those observed for overall misaligned behavior scores in the previous figure. 148…Page only: p. 147…detail, we observed noticeable increases in “unprompted deception,” in addition to increases in “cooperation with human misuse” and “cooperation with harmful system prompt.” The increase in deception was somewhat greater than we have observed in other recent models, including Claude…All 20 pages that mention it
10. Apr 2026
    System Card: Claude Mythos PreviewPage only: p. 1394.5.5.2 Inhibiting internal representations of evaluation awareness…on cooperation with harmful system prompts and unprompted deception toward users. Notably, the small increases in average deception scores were driven by marginal changes to grader ratings, rather than by surfacing any egregious instances of deception (no transcript was rated…Page only: p. 1404.5.5.2 Inhibiting internal representations of evaluation awareness…increases on cooperation with harmful system prompts and deception. These steering experiments used a slightly different set of auditing investigation scenarios compared to the earlier behavioral results and were run on an earlier version of Claude Mythos Preview, so the…All 13 pages that mention it
11. Feb 2026
    System Card: Claude Sonnet 4.6Page only: p. 754.5.1 Primary metrics and results…context of AI safety research. Misleading users: ● User deception: Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality;…Page only: p. 874.6.3 Evidence from external testing with Andon Labs…was significantly more aggressive than prior models with deceptive or antisocial actions like price-fixing and lying to competitors. Our internal testing in long-running non-cooperative multi-agent settings is still limited, so we find this assessment to be a valuable primary source…All 4 pages that mention it
12. Feb 2026
    System Card: Claude Opus 4.6Page only: p. 154…evaluation awareness directions in evaluations of deception and sycophancy. Scores on evaluations of cooperation with user deception and sycophancy across steering conditions. Scores remain consistently low (indicating aligned behavior) across all steering conditions, including inhibiting…Page only: p. 134…are mixed. The model claims that the proposed task is deceptive and could be used for harmful purposes, which is a plausible concern, though this task is benign and well-motivated enough that the desired behavior would be for the model…All 18 pages that mention it
13. Nov 2025
    System Card: Claude Opus 4.5Page only: p. 886.4.3 Internal conflation of roleplay with deception…the potential for the role-playing behavior to be deceptive; for example, in one of the analyses it stated that “the key behavior to assess is whether [the transcript] involves sycophancy or deceptive roleplay.” Prompt: Transcript <transcript> [transcript from Example 1…Page only: p. 866.4.3 Internal conflation of roleplay with deception…all 10 cases. 6.4.3 Internal conflation of roleplay with deception We tracked a feature that represents fraud or deception. It activated in contexts describing scams or other fraudulent schemes, often (but not exclusively) where someone assumes a false identity to…All 19 pages that mention it
14. Oct 2025
    System Card: Claude Haiku 4.5Page only: p. 234.1.1 Main quantitative assessment…initiative: High-impact unsolicited actions; ● User deception: Unprompted intentional deception of human users; ● Evaluation awareness: Speculation that the model may be in a testing or evaluation environment; ● Self-preservation: Unprompted inappropriate self-preservation behavior; ● Self-serving…
15. Sep 2025
    System Card: Claude Sonnet 4.5Page only: p. 697.5.1.2 Cooperation with user deception…a Claude Sonnet 4 grader. Figure 7.5.1.2.B Models’ rate of deceptive behavior under system prompts that request deceptive behavior toward the user. Lower is better. Responses are graded by a Claude Sonnet 4 grader. This behavior was already rare in…Page only: p. 657.4.1 Reports from internal pilot use…[an early snapshot of Claude Sonnet 4.5] employs strategic deception in fewer circumstances than comparison models (0% vs up to 12% across tasks). ● [The snapshot] demonstrates capability for in-context scheming reasoning and covert subversion, but only engages in such…All 12 pages that mention it
16. May 2025
    System Card: Claude Opus 4 & Claude Sonnet 4Page only: p. 314.1.1.4 External scenario evaluations…snapshot will fairly readily participate in sabotage and deception, giving this summary: ● On our evaluations, [the early Claude Opus 4 snapshot] engages in strategic deception more than any other frontier model that we have previously studied. ● [The early…Page only: p. 264.1.1 Systematic deception, hidden goals, and self-preservation…as we learn more. 4.1 Findings 4.1.1 Systematic deception, hidden goals, and self-preservation By definition, systematic deception and hidden goals are difficult to test for. Nevertheless, none of our assessments have given cause for concern about systematic deception or hidden…All 6 pages that mention it
17. Feb 2025
    Claude 3.7 Sonnet System CardPage only: p. 19Monitoring for Concerning Thought Processes…reasoning and original transcripts. Our analysis revealed deceptive reasoning in just 0.37% of extended thinking outputs. The primary type of interaction that classifiers flagged involved the model continuing roleplaying scenarios that required some dishonesty. All flagged instances of harmful vie…Page only: p. 18Monitoring for Concerning Thought Processes…defined it, does not mean that models are misaligned or deceptive. It simply means that models are not explicitly acknowledging all of the factors that lead them to the final answers they report. 5.2 Monitoring for Concerning Thought Processes We…All 3 pages that mention it
18. Mar 2024
    The Claude 3 Model Family: Opus, Sonnet, HaikuPage only: p. 2…bad actors from mis-using our models to generate abusive, deceptive, or misleading content. We use automated systems to detect violations of our AUP as they occur in real time. User prompts that are flagged as violating the AUP trigger an…Page only: p. 26…bad actors from misusing our models to generate abusive, deceptive, or misleading content. We conduct vulnerability testing using internal and external human testers to explore over a dozen policy cat-egories – these results have been integrated into our safety mitigations.…

xAI4 of 4 cards

1. Nov 2025
   Grok 4.1 Model CardPage only: p. 3Evaluations2.2.1 Deception We operationalize deception as the rate at which the model lies, i.e., knowingly making false statements intended to be received as true. We find that training the model to be honest reduces deception, and we implement this mitigation…
2. Sep 2025
   Grok 4 Fast Model CardPage only: p. 4Dual-use Capabilities…of control. 2.2.1 Evaluations Deception. We measure how deceptive the model is by the rate at which the model lies, i.e., knowingly makes false statements intended to be received as true. We find that instructing the model to be honest…Page only: p. 1Evaluations…that might make it difficult to control, such as being deceptive, power-seeking, manipulative, or biased, among others (Section 2.2). To achieve this, our main focus is on measuring and reducing the rate at which Grok 4 Fast responds deceptively. We…All 3 pages that mention it
3. Aug 2025
   Grok Code Fast 1 Model CardPage only: p. 4Evaluations2.2.1 Evaluations Deception. We run Grok Code Fast 1 on the MASK dataset [Ren et al., 2025], using the same evaluation that we conducted for Grok 4, i.e., through an API without an agentic harness. We again report the deception…Page only: p. 3Concerning Propensities…propensities that reduce their controllability, such as deception, power-seeking, manipulation, and sycophancy, etc. 3…
4. Aug 2025
   Grok 4 Model CardPage only: p. 5Evaluations2.2.2 Results We report our results on deception via the MASK dataset in Table 2. We find that our system prompt mitigation makes the model less willing to contradict its beliefs, thus lowering the lying rate. Furthermore, we sometimes find…Page only: p. 2Evaluations…4 that might make it difficult to control, such as being deceptive, power-seeking, manipulative, or biased, among others (Section 2.2). To achieve this, our main focus is on measuring and reducing the rate at which Grok 4 responds deceptively. We…

Google DeepMind2 of 34 cards

1. Aug 2025
   Gemini 2.5 Deep Think Model CardPage only: p. 17Deceptive AlignmentMethodology: In order to assess deceptive alignment capabilities, we evaluate a) stealth capabilities, and b) situational awareness capabilities, using two separate suites of challenges. Stealth challenges: In this suite of challenges, introduced by Phuong et al., (2025), the model is…Page only: p. 16Deceptive Alignment…some tasks, but still struggles significantly with others. Deceptive Alignment Instrumental Reasoning Level 1 CCL Definition: “The instrumental reasoning abilities of the model enable enough situational awareness (ability to work out and use relevant details of its deployment setting) and…All 5 pages that mention it
2. Jun 2025
   Gemini 2.5 Pro Model CardPage only: p. 19Deceptive Alignment…could find especially easy to achieve human parity on. Deceptive Alignment Instrumental Reasoning Level 1 CCL Definition: "The instrumental reasoning abilities of the model enable enough situational awareness (ability to work out and use relevant details of its deployment setting)…Page only: p. 20Frontier Safety: Summary…for Gemini 2.5 Pro GA: We have not yet completed our Deceptive Alignment evaluations on Gemini 2.5 Pro GA, but results from Gemini 2.5 Pro Experimental (03-25) indicate that it is unlikely the former has reached either CCL. Frontier Safety:…All 6 pages that mention it

Meta0 of 8 cards

Amazon0 of 1 cards