Anthropic introduced self-assessment on Oct. 16, 2026, to have Claude assess conversations for potentially harmful requests after it responds. The company says a short follow-up instruction prompts Claude to produce a preliminary assessment covering the possible harm, its type, a priority for human review, whether Claude refused and a brief explanation. Those assessments feed additional offline review, including for conversations in which Claude refused a request.
Anthropic says the monitor runs across a large fraction of Claude’s global traffic. In an Oct. 29 post, the company said it had used self-assessment since June and that it surfaced misuse other safeguards were not designed to catch. Anthropic estimates each assessment generates about 10 additional output tokens on average and adds less than 2% to inference compute across its production API traffic distribution.
What the performance numbers measure
For Claude Fable 5, Anthropic reports that self-assessment flags 78% of conversations its stronger prompted classifier flags as potentially harmful and 94% of that classifier’s high-priority conversations, at a 1% false-positive rate. The benchmark is Anthropic’s own prompted classifier, which receives the conversation and the company’s safeguards policies, not an independent determination of harmful use.
Anthropic says self-assessment does not catch every harmful conversation and is not robust enough to stop a dedicated attacker. It describes the system as an offline monitoring layer alongside separate safeguards that block harmful requests in real time.
Comments