Anthropic's AI Is Too Cautious (And Security Researchers Are Furious)

Anthropic launched Fable, an AI model with a noble idea: protect the world from bad actors by adding guardrails—basically limits and controls that prevent the AI from doing potentially dangerous things. Sounds good on paper. The catch? These guardrails are so tight that security experts can't even use the model for their legitimate day job.

The researchers complaining aren't trying to hack other people's systems—they want to test their own systems to find holes before actual hackers do. It's the most boring and necessary job ever, but Fable says no thanks. Penetration testing, or deliberately searching for vulnerabilities in a controlled way, is standard industry practice. Not villain stuff.

Here's the core tension: when you lock down an AI so much that you block even the "good guys" from using it, are you solving a problem or creating a new one? Because if security experts can't use modern models for their defensive work, they'll either dust off old tools, get less sophisticated help, or—here's the irony—start hunting for models with fewer restrictions from other companies.

Anthropic actually has thoughtful ideas about AI safety, but the practical result feels like an airbag that deploys every time you brake gently. Protective? Yes. Frustrating? Also yes.

Right now it's an undercurrent tension between wanting to shield humanity from doomsday scenarios and the real-world need to make our current defense tools actually work. It's not a problem a software update fixes.