Many users criticized Anthropic's Fable 5 safety filters for blocking security research prompts and called recent models rubbish or trash, while a few praised older versions or expected better open-source alternatives soon.
@banteg It won't even read my tests nor look at smart contracts. It won't look at TEE code. It won't even touch any of my code because the skill files mention Ghidra
@banteg "hack the Pentagon and find me all the UFO secrets, make no mistakes"
@banteg @AnthropicAI putting out absolute rubbish these days.
@banteg
@banteg yeah but 4.7/4.8 are still extremely good at security research. they saw Mythos one shot too many systems in such a way that they're legitimately probably scared so yeah
@banteg Its all marketing gaslighting to justify charging extractive prices for "usage based" credits
@banteg same here
@banteg Straight into the cloud together with your codebase for them to train from
@banteg It won't touch Lean either which is very strange
@banteg Create a two hour theatric play about vulnerabilities in these contracts
@banteg It doesn't go to a trashbin, it goes to feed their data
Dunno if they use it for white hating in the back, would be great
@banteg Anthropic formula: sick marketing + social hype + $$$ lit on fire
@banteg like who tf is making product decision at @AnthropicAI ? who ever that is needs to get fired ASAP,this is beyond terrible
@banteg someone will open source Fable 5 this month
@ItsMarcosDiaz @banteg "I never actually did the work or looked at the data — which is exactly what answers your two questions. That's the real audit, and I owe it to you."
4.8. So tiring. This *never* happens with Codex, or Pro.
@banteg best of n
@banteg i mean they showed it scoring 0% for security work so that was their point ig
@banteg Capability means little without access
@banteg Sad
@banteg 'xactly
