Anthropic's Fable Guardrails Are Backfiring Badly

Two separate threads are tearing into Anthropic's new Fable model from different angles. The first is about the 30-day mandatory data retention policy, which applies even to enterprise customers using AWS Bedrock. One commenter flagged that the 'almost all cases' language in Anthropic's deletion policy is doing a lot of heavy lifting, meaning companies can't actually guarantee their data disappears. The second thread covers cybersecurity researchers who say Fable's content filters block legitimate vulnerability research while a determined attacker just rewrites the prompt and gets through anyway. Worse, commenters allege the model silently switches to a weaker version for flagged topics without telling the user it has done so.

The pattern here: Anthropic is trying to thread a needle between safety and utility, and it's cutting itself on both ends. Enterprise customers can't trust the privacy guarantees, and technical users can't rely on the model to behave predictably. DeepSeek is getting name-dropped in the security thread as the model that actually helps with PoC vulnerability work, which is not a comparison Anthropic wants.

The counterpoint is simple: 'Then don't use it.' But that dismissal misses the real issue. When the market leader in 'responsible AI' has guardrails that sophisticated users route around while blocking legitimate researchers, the safety theater argument gains a lot of traction. Trust is the product for an AI company, and both threads are eroding it.