Anthropic has added “self-assessment,” a background safety check that reviews Claude conversations after the model responds, complementing—not replacing—real-time safeguards.

Running since June 2026 across a large share of Claude’s global traffic, the feature is available on every model released since the Fable 5 wave.

The check identifies potentially harmful conversations, categorizes the possible harm, assigns a priority for human review, and can note whether Claude refused the request; it runs silently, so safe chats continue without warnings or interruptions.

Flagged conversations go to human reviewers for offline investigation rather than being blocked immediately, while ordinary safe conversations pass without user-facing alerts.

That makes self-assessment an added layer in Anthropic’s defense-in-depth strategy: real-time controls remain responsible for stopping harmful requests before responses, while the later review can surface attacks that slip through.

On Claude Fable 5, Anthropic says the system flags 78% of potentially harmful conversations and identifies 94% of high-priority cases, with a roughly 1% false-positive rate; some harmful activity still goes undetected.

The feature is not a standalone defense: in cybersecurity tests, attackers bypassed it 54% of the time, compared with 79% when there were no safeguards and 25% when a real-time blocking system was used.

Anthropic cautions that the system can miss harmful activity, especially when users deliberately try to evade safeguards.

Anthropic said that in August 2026 it flagged more than 100,000 cases of child exploitation material associated with one customer and reported them to authorities.

The company says the feature does not collect additional user data, though flagged conversations may be retained for investigation under its data-handling policies.

Anthropic says each assessment adds about 10 tokens and less than 2% to computing costs, compared with nearly 24% for its older safety systems.

The rollout follows Anthropic’s October 2026 expansion of access for third-party evaluators, including METR and SecureBio, as part of broader efforts to monitor and disclose potential model misuse.