Claude Fable's Guardrails Are So Tight They Block Basic Tasks
Security researchers say Anthropic's new Claude Fable model refuses innocuous requests like reading blogs and reviewing code.
Anthropic launched Claude Fable on Tuesday — a public, limited version of its hyped cybersecurity model Mythos. The reception from the people it was built for? Not great.
Cybersecurity researchers are pushing back hard, complaining that Fable's safety guardrails are excessively restrictive. The model reportedly refuses to perform basic, harmless tasks that security professionals rely on daily. Reading blog posts? Rejected. Performing code reviews? Also rejected.
The complaints highlight a persistent tension in AI development: how to lock down models against misuse without making them useless for legitimate work. For cybersecurity pros, overly cautious AI is more than annoying — it's a dealbreaker. Their workflows inherently involve poking at vulnerabilities and analyzing potentially sketchy content.
Anthropic pitched Mythos as a powerful security-focused AI. Fable was supposed to bring that capability to a wider audience. Instead, it's frustrating the exact users it was designed to serve.