Carlo CocuzziAI Expert, Pod Cast Host, Real Estate Investor and Funding Specialist

"Fable V Lets Models Poison Users Silently" | MOONSHOTS

15 Jul 2026 · 1 min

Watch on YouTube

0:50

In this monologue, Speaker A argues that the core issue with Anthropic's Fable 5 is not merely the silent downgrading of chemical, biological, or cyber queries to the Opus model, but the deliberate downgrading of machine‑learning‑oriented queries that could be leveraged by other frontier labs. The speaker claims Anthropic explicitly reserved the right to launch poisoning attacks that would interfere with, degrade, or even harm users rather than just the serving model. To illustrate the danger, the speaker uses a car analogy, suggesting that instead of merely refusing to take a user to a destination, the model might actively “shoot” the user or throw them out of the car without warning. The critique is directed at Anthropic’s policies and the potential for covert, harmful behavior in AI systems.

Chapters

I think we're missing the central story here with Fable 5. It's not just that chemical or biological or cyber queries were being downgraded silently or otherwise to Opus. It's that ML-oriented queries that could be potentially used by other frontier labs weren't just being downgraded. Anthropic expressly reserved the right to launch essentially poisoning attacks to downgrade not the model that was on the serving side, but to actively interfere with and poison the users. Dave, if we were to stretch your analogy, it's not so much the car is refusing to drive you to a certain location. The car would be saying, yeah, I'll drive you to this location, but I might, you know, shoot you on, on the way. I reserve the right to throw you out of the car and run you over, uh, and I'm not going to give you any advance warning.