Anthropic has unveiled its latest model, Fable 5, with some unusual restrictions. The company is keeping it from discussing topics like cybersecurity and biology, fearing that malicious actors might misuse this information.

The new safeguards are strict. If you ask Fable 5 about certain sensitive areas, it will direct your query to an older version of its model instead—a bit like a parental filter on the internet, but for AI. The company hopes these measures prevent serious harm from falling into the wrong hands.

Anthropic’s worries stem from previous versions that could potentially assist in “agentic hacking.” In tests, even with these new limits, Fable 5 still held up better against automated attacks than its predecessors. It’s a step forward, but not an unqualified victory.

Their approach is to err on the side of caution, knowing that sometimes it might be frustrating for users who just want to know things. But in a world where AI can do real damage, maybe that’s better than risking everything?