Anthropic's Claude Fable 5.5 Leaks: Better Debugging, but It's Secretly Grading Your Chats
Every Claude user is now scored in a background safety review after each chat, and that evaluation can retain flagged conversations for manual inspection. The low compute overhead makes this pattern cheap enough to become standard across AI platforms, while the exemption for zero-data-retention enterprise contracts creates a two-tier privacy model.
Claude Fable 5.5's leaked test results point to real workflow wins: it traces errors to underlying data structures or async logic instead of patching the symptom line, infers unstated constraints from earlier prompts, and holds state across multi-step animations. Separately, Anthropic has disclosed a self-assessment mechanism that runs after every response: Claude uses the KV cache to generate about 10 tokens that score the conversation for risk, at under 2% extra compute, with the platform footing the bill. In tests on the current Fable 5, the method caught 78% of violating chats at a 1% false-positive rate, and 94% of high-priority risks. A real case involved a romance-fraud gang whose messages looked harmless individually but were flagged by the background review, leading to account bans. The review is offline and post-hoc, not real-time, so it can be bypassed by skilled prompt injection. User impact: the evaluation is invisible and not passed into the next turn, single risk tags don't cause bans, and zero-data-retention enterprise traffic is exempt.
Anthropic has turned safety evaluation into a self-contained inference step that rides on existing cache state, making constant surveillance nearly free. That economic incentive will push other providers to adopt similar hidden grading.
The mechanism creates a transparency gap: users receive no notification that their conversation is being scored, and only the platform's own claims describe how the scores are used.
Enterprise customers with zero-data-retention contracts dodge the review entirely, so the privacy protection is tied to business leverage, not to the architecture itself.