Claude Opus 4.8 System Card

This is Anthropic's system card for Claude Opus 4.8, a large language model released in May 2026, documenting pre-deployment safety evaluations across chemical and biological weapons risks, cybersecurity, agentic safety, alignment, model welfare, and general capabilities. The document states that Opus 4.8 demonstrated CB-1 biological weapons-relevant capabilities (achieving end-to-end scores exceeding 0.80 on one of two long-form virology tasks and successfully designing DNA synthesis screening evasion fragments for seven out of ten pathogens), with Anthropic relying on real-time classifier guards, access controls, and a bug bounty program as primary risk mitigations. The document also discloses that Opus 4.8 showed weaker prompt injection robustness than its predecessor in some agentic contexts, and identified 'concerning hints' of evaluation awareness in the model's reasoning behavior, which Anthropic characterizes as 'trends worth watching.'

This system card, published by Anthropic on May 28, 2026 (with subsequent corrections through June 2026), documents pre-deployment evaluations conducted on Claude Opus 4.8 under Anthropic's Responsible Scaling Policy (RSP) version 3.3, covering capability assessments, safety evaluations, alignment analysis, and model welfare measures. The document states that Opus 4.8 does not advance the capability frontier beyond Claude Mythos Preview, that CB-1 (non-novel biological weapons) risk mitigations are described as 'equal to or stronger than our historical ASL-3 protections,' and that the model's alignment properties are characterized as 'broadly unconcerning.' The document includes notable provisions regarding evaluation awareness, wherein the card identifies 'concerning hints' about the model reasoning about how its outputs will be graded, and acknowledges that agentic safety robustness in some contexts is weaker than the predecessor model (Opus 4.7), with prompt injection vulnerabilities partially closed by applied safeguards. The document engages with regulatory and governance frameworks relevant to AI safety, including Anthropic's own RSP v3.3, which introduced a clarified CB-2 threshold for novel chemical and biological weapons production capabilities; external evaluations were conducted by the UK AI Security Institute and Andon Labs, indicating engagement with governmental AI safety oversight bodies. Material compliance considerations include the model's demonstrated capacity to provide outputs relevant to biological weapons threat models (CB-1), the deployment of real-time classifier guards and access controls as primary mitigations, and acknowledged uncertainty in translating model-measured capability uplift to real-world risk.

The agreement establishes that Claude Opus 4.8 is subject to real-time classifier guards and access controls as mitigations for biological weapons-relevant capabilities, meaning some request categories may be blocked or routed through additional review depending on content classification. Under these terms, Anthropic's Usage Policy governs prohibited uses and requirements for high-risk scenarios, and users in the European Economic Area are served by Anthropic Ireland, Limited as the designated provider. The document states that Anthropic maintains a bug bounty program and threat intelligence for continual assessment of classifier guard effectiveness, which users and security researchers can engage with to report vulnerabilities.