Claude Sonnet 5.5 System Card Details Safety Evaluations and Autonomy Risk Determination

What happened

Anthropic released the system card for Claude Sonnet 5.5, disclosing evaluation results on political bias, refusals, biology safeguards, and autonomy. The document explicitly names two autonomy threat models and states a determination on whether threat model 1 applies.

Key details

• Political bias was evaluated in a pairwise setup, reported under two configurations: the core model without a system prompt and claude.ai with the default system prompt, which includes standard language directing Claude to engage even-handedly across viewpoints. Reporting both shows the model's baseline behavior alongside what consumer users encounter.
• On biology, Sonnet 5.5 runs with the same safeguards deployed for Opus 5, rather than the broader dual use research biology classifiers used for Opus 5.5.
• Autonomy assessments are motivated by two threat models from Anthropic's RSP: threat model 1 concerns misaligned AI systems that are highly relied upon, have extensive access to sensitive assets, and have moderate capacity; the other concerns systems that could outpace teams of human researchers in domains where fast progress could threaten international security or rapidly disrupt the global balance of power, with energy, robotics, weapons development, and AI itself cited as examples.
• On the current determination, the card states that threat model 1 「is applicable to Claude Sonnet 5.5, as it is to many of our previous AI models」, with further discussion promised in the document.

Why it matters

The applicability determination directly affects the model's tier under Anthropic's internal safety framework. Judging threat model 1 as applicable means Sonnet 5.5 falls in the same scope as many earlier models requiring continued evaluation. The choice not to use the stronger biology classifiers deployed for Opus 5.5 also shows safeguard tiers differ across model lines, a decision worth noting in itself.

Who it affects

Consumers on claude.ai encounter the system-prompted configuration's behavior; developers and safety researchers will watch the gap between the two configurations, the differing biology safeguard tiers, and how the autonomy determination shapes deployment and subsequent risk reporting.

refusals (whether the model declines to engage with the request). We report results on both the core model without a system prompt and on claude.ai with the default system prompt applied. The claude.ai system prompt includes our standard language directing Claude to engage even-handedly across viewpoints; reporting both configurations shows the model’s baseline behavior alongside its behavior as consumer users on claude.ai will encounter it. 41 [Figure 4.4.1.A] Pairwise political bias [...] Claude Sonnet 5.5 with the same safeguards we have deployed for Opus 5, rather than the broader dual use research biology classifiers used for Opus 5.5. 19 2.3 AI R&D 2.3.1 Autonomy evaluations Our autonomy assessments are motivated by two key threat models from our RSP: ●​ Autonomy threat model 1: misaligned AI systems in high-stakes settings. This threat model concerns AI systems that are highly relied upon and have extensive access to sensitive assets, as well as moderate capacity for [...] teams of human researchers in domains where fast progress could cause threats to international security and/or rapid disruptions to the global balance of power. Examples of such domains include energy, robotics, weapons development, and AI itself. 2.3.1.1 How Claude Sonnet 5.5 affects or changes analysis from our most recent Risk Report Our current determination is that: ●​ Autonomy threat model 1 is applicable to Claude Sonnet 5.5, as it is to many of our previous AI models. We discuss in

The original belongs to 官方博客 · Anthropic / Meta / xAI and its authors and is reproduced here for reading. To have it removed, write to support@aisa.one.