Anthropic Deploys Post-Response AI Self-Assessment to Detect Misuse
Emerging Technology Risk · 3 sources · Updated 19d ago · First seen Dec 8, 2025

Anthropic reports deploying self-assessment, an offline monitoring technique in which Claude reviews conversations for potential misuse after responding. On Claude Fable 5, the technique detected 78% of conversations flagged by a stronger prompted classifier and 94% of high-priority conversations at a 1% false-positive rate, while adding less than 2% to inference compute. Anthropic says production deployment helped uncover previously undetected abuse, prompt injection attempts, unauthorized agent actions, and an automated romance-scam operation. These results measure agreement with a reference classifier, rather than detection of all actual misuse.
The approach complements research into both preventive safeguards and post-response accountability. Constitutional Classifiers, trained using synthetic data derived from natural-language policies, reportedly resisted universal jailbreaks across more than 3,000 estimated hours of red teaming; enhanced classifiers added 23.7% inference overhead and increased production-traffic refusals by 0.38 percentage points. Separately, Training LLMs for Honesty via Confessions describes training GPT-5-Thinking to disclose policy violations after answering, with confession rewards based solely on honesty and separated from the original answer’s reward. For security teams, these approaches support layered defenses rather than reliance on model self-reporting: Anthropic warns that self-assessment remains vulnerable to targeted evasion, inherits its teacher classifier’s blind spots, and cannot replace real-time safeguards or human investigation.


