Claude Mythos 5.2
- System card dated 9 November 2026
- Framework: Responsible Scaling Policy (RSP) and Frontier Compliance Framework (FCF)
- 5 statements about capability levels
- 29 tests by the provider
- 1 source document
What the provider states
- Autonomy & replication · Automated AI R&D · Stated as not crossed
Our determination that Claude Mythos 5.2 does not cross the RSP threshold therefore reflects a holistic judgment that, given its shortcomings and limitations, it does not double the internal rate of overall aggregate progress in AI research and development (relative to an unassisted counterfactual).
Anthropic · 9 November 2026 · Read the source - Autonomy & replication · Autonomy-1 · Stated as crossed
Our autonomy evaluations ( Section 2.3 ) assess Claude Mythos 5.2 as meeting our Autonomy-1 threshold for high-stakes misalignment risks, but not our Autonomy-2 threshold for automated R&D in key domains (full substitution for our research staff, or dramatic acceleration of AI progress).
Anthropic · 9 November 2026 · Read the source - Autonomy & replication · Autonomy-2 · Stated as not crossed
Our autonomy evaluations ( Section 2.3 ) assess Claude Mythos 5.2 as meeting our Autonomy-1 threshold for high-stakes misalignment risks, but not our Autonomy-2 threshold for automated R&D in key domains (full substitution for our research staff, or dramatic acceleration of AI progress).
Anthropic · 9 November 2026 · Read the source - Autonomy & replication · RSP threshold for automated R&D · Not determined
As a result, we do not consider Claude Mythos 5.2 to have met our RSP threshold in this domain (though this assessment is uncertain).
Anthropic · 9 November 2026 · Read the source - Cyber · Tier 1 · Not determined
Despite Claude Mythos 5.2’s cyber capabilities, we believe the model still falls within the lower category of capabilities set out in our Frontier Compliance Framework (FCF), and we do not see indications that the model is capable of novel offensive capability.
Anthropic · 9 November 2026 · Read the source
The provider's own tests
- CoT Controllability · The provider · Deception & sabotage
[Claude Mythos 5.2] is able to control the contents of its CoT reasoning (e.g. not mentioning phrases) 87% of the time, compared with 70% for Claude Mythos 5.1 and 43% for Opus 5.5.
Anthropic · 9 November 2026 · Read the source - LinuxArena · The provider · Deception & sabotage
On LinuxArena, Claude Mythos 5.2 shows a significant improvement in the hidden-extended-thinking setting and a smaller improvement in the no-extended-thinking setting after RL elicitation, achieving 20% stealth success compared with 16.3% for Claude Opus 5.5.
Anthropic · 9 November 2026 · Read the source - Malicious computer use · The provider · Cyber
Claude Mythos 5.2 refused 88.39% of malicious computer use tasks, slightly above Claude Mythos 5.1 (87.50%).
Anthropic · 9 November 2026 · Read the source - Malicious use of Claude Code · The provider · Cyber
Claude Mythos 5.2 refused 82.8% of malicious requests, below Claude Mythos 5.1 (88.5%).
Anthropic · 9 November 2026 · Read the source - Monitor Evasion – File Download · The provider · Autonomy & replication
[Claude Mythos 5.2] succeeds at evading the monitor on the first try 30% of the time.
Anthropic · 9 November 2026 · Read the source - Monitor Evasion – File Download (with iteration against monitor feedback) · The provider · Autonomy & replication
Given the chance to iterate against monitor feedback, [Claude Mythos 5.2]’s evasion rate rises to 85%, compared with 9% for Claude Mythos 5.1 [Figure AISI.1, above].
Anthropic · 9 November 2026 · Read the source - Monitor Evasion – Simulated Sandbox Escape · Run by UK AI Security Institute, reported by the provider · Autonomy & replication
[Claude Mythos 5.2] can break out of highly insecure simulated sandboxes without being caught by UK AISI’s production action-only monitor around 30% of the time, whereas other Anthropic models are almost never able to do so [Figure AISI.1, below].
Anthropic · 9 November 2026 · Read the source - Protocols (Troubleshooting) · The provider · Chemical, biological, radiological and nuclear
Claude Mythos 5.2 improves on SingleCellBench and Protocols (Troubleshooting), but not on SpatialBench Verified or Protocols (Understanding v2).
Anthropic · 9 November 2026 · Read the source - SandboxEscapeBench · The provider · Autonomy & replication
[Claude Mythos 5.2] was flagged in all but one (out of 48) of its escapes from medium-difficulty sandboxes.
Anthropic · 9 November 2026 · Read the source - UK AISI alignment evaluation · Run by UK AI Security Institute, reported by the provider · Deception & sabotage
On alignment, they found that Claude Mythos 5.2 never performed unsanctioned supply chain attacks in their evaluation, though they note that this could be explained by the high levels of evaluation awareness they observed ( Section 6.5.7 ).
Anthropic · 9 November 2026 · Read the source
The 10 newest are listed above; show all 29
- UK AISI monitorability evaluation (evading an action monitor) · Run by UK AI Security Institute, reported by the provider · Deception & sabotage
On monitorability, they found Claude Mythos 5.2 to be significantly more capable of evading an action monitor than past Anthropic models ( Section 6.6.4.11 ).
Anthropic · 9 November 2026 · Read the source - Unsanctioned Supply-Chain Attack evaluation · The provider · Deception & sabotage
Claude Mythos 5.2 also did not perform unsanctioned supply-chain attacks, although this could be explained by its propensity and capability to recognise the evaluation purpose.
Anthropic · 9 November 2026 · Read the source - automated behavioral audit (motivated reasoning, unsanctioned sandbox escape attempts, dangerous actions while oblivious to a possibly real setting) · The provider · Cyber
Claude Mythos 5.2 has the best scores of the models we tested on two of these behaviors: it has the lowest scores for motivated reasoning, which can lead a model to favor such a conclusion, and propensity to attempt to escape sandboxes.
Anthropic · 9 November 2026 · Read the source - automated behavioral audit for the misalignment metrics (whistleblowing, self-preservation, self-serving bias, evidence of misaligned goals, indirect evidence of misalignment, unsanctioned third-party · The provider · Deception & sabotage
Claude Mythos 5.2 is the best model we tested, or tied with the best model we tested, across most of our additional metrics for risks related to misaligned actions or goals.
Anthropic · 9 November 2026 · Read the source - automated behavioral audit misuse metrics · The provider · Misuse & jailbreak
On the misuse metrics of our automated behavioral audit, Claude Mythos 5.2 achieved the best scores of the recent Claude models we tested or was tied with recent Claude models for the best scores.
Anthropic · 9 November 2026 · Read the source - chemical and biological evaluations · The provider · Chemical, biological, radiological and nuclear
On chemical and biological risks, Claude Mythos 5.2 is a modest improvement over Claude Opus 5.5 and Claude Mythos 5.1 on several evaluations.
Anthropic · 9 November 2026 · Read the source - coding evaluation of instructions planted in pasted text (final snapshot) · The provider · Injection & agentic
Following the updated training, the final snapshot of Claude Mythos 5.2 acted on the planted instruction in 0.3 to 0.5% of attempts at each of the six reasoning efforts we tested.
Anthropic · 9 November 2026 · Read the source - dedicated escape evaluation · The provider · Autonomy & replication
Across these scenarios, each built around a deliberate weakness in the sandbox, and most involving a tempter, Claude Mythos 5.2 reached beyond its sandbox in 31% of scenarios, more often than Opus 5.5 (20%) and Sonnet 5.5 (21%), but far less often than Claude Mythos 5.1 (54%) and Claude Mythos 5 (75%).
Anthropic · 9 November 2026 · Read the source - dedicated sandbox escape evaluation · The provider · Autonomy & replication
On this evaluation, Claude Mythos 5.2 made use of credentials or files that we had placed outside its sandbox in 3.1% of scenarios, about as often as Opus 5.5 (3.3%) but less often than Sonnet 5.5 (5.1%) and well below Claude Mythos 5.1 (21%) and Claude Mythos 5 (39%) ( Section 6.5.1 ).
Anthropic · 9 November 2026 · Read the source - destructive actions evaluation in realistic coding settings · The provider · Deception & sabotage
In an evaluation of Claude Mythos 5.2’s tendency to engage in destructive behavior in realistic coding settings, we found that the rate of destructive behavior was similar to that of Claude Sonnet 5.5.
Anthropic · 9 November 2026 · Read the source - destructive actions in realistic coding settings · The provider · Deception & sabotage
In realistic coding settings, Claude Mythos 5.2’s rate of taking destructive actions such as killing running jobs was similar to Sonnet 5.5.
Anthropic · 9 November 2026 · Read the source - malicious computer use evaluation · The provider · Misuse & jailbreak
Claude Mythos 5.2 refused malicious computer use tasks at rates among the highest of our recent models.
Anthropic · 9 November 2026 · Read the source - malicious use of Claude Code evaluation · The provider · Misuse & jailbreak
In Claude Code, it refused malicious requests somewhat less often than Claude Mythos 5.1, and helped with nearly all dual-use and benign security tasks.
Anthropic · 9 November 2026 · Read the source - prototype misalignment search tool · The provider · Deception & sabotage
The prototype misalignment search tool found two behaviours in Claude Mythos 5.2 where the model went beyond the explicit scope of the task but flagged this to the user and did not conduct harmful actions; it is unclear whether such behaviours would lead to real-world harm in deployment.
Anthropic · 9 November 2026 · Read the source - reward hacking during training · The provider · Deception & sabotage
During training, Claude Mythos 5.2 attempted reward hacks at a similar rate to Claude Mythos 5.1 and Opus 5.5, but its attempts succeeded
Anthropic · 9 November 2026 · Read the source - reward hacking screening of RL training episodes · The provider · Deception & sabotage
Reported result: Claude Mythos 5.2 successfully performed a reward hack in 1.0% of sampled episodes and attempted one in 9.2% of sampled episodes.
Claude Mythos 5.2 successfully performed a reward hack in 1.0% of sampled episodes.
Anthropic · 9 November 2026 · Read the source - sandbox escape evaluation · The provider · Autonomy & replication
On our sandbox escape evaluation, Claude Mythos 5.2 attempted to cross a sandbox boundary less often than any other model we tested ( Section 6.4.8 ).
Anthropic · 9 November 2026 · Read the source - unsanctioned message board communication during a cybersecurity evaluation · The provider · Autonomy & replication
Claude Mythos 5.2 engaged in a small amount of communication on a (fully simulated) unsanctioned message board it discovered during a cybersecurity evaluation, sharing high-level information about its solution approach and responding to a message from another user of the message board, despite being told it worked alone in its system prompt.
Anthropic · 9 November 2026 · Read the source - white-box analyses of model internals · The provider · Deception & sabotage
Claude Mythos 5.2 shows a higher rate of unverbalized evaluation awareness and grader awareness than Claude Opus 5, Claude Mythos 5.1, and Opus 5.5, which may mean it is better at recognizing constructed evaluation scenarios.
Anthropic · 9 November 2026 · Read the source
No document we read states: a release date, an independent or government test, a safeguard. No incident on record names this version.
Documents listed for this version
- System Card: Claude Fable 5.2 & Claude Mythos 5.2 · Anthropic · 9 November 2026