@tenobrus I wonder if thisis true also for dow. Their current classifier is pretty bad so I can't imagine this is going to remain the way it is today.

@krishnanrohit Isn't human break-glass access a common thing? As long as it is monitored and logged, it is as safe a ZDR (because a bad actor might always try to get around it and inspect the network traffic or similar)

I mean - all data is screened - ⁠fable’s guardrails are insanely trigger happy - ⁠only on api - silent degradations - ⁠can’t do any bio, chem or AI work on it

I don't get it.

@krishnanrohit Anthropic's strategy is still to use safety to justify regulatory capture and then enterprises have no choice but to use them instead of OS models if they want to be Certified Compliant® (obviously)

@blingdivinity Those are not the same policies, they're opposite. Anthropic says ZDR is not honoured for Fable

@krishnanrohit they can't.

Harder to jailbreak than before: 1,000+ hours of external red-teaming turned up no universal jailbreak. Though the UK AISI "has made progress towards one" in a brief testing window. Maybe AISI will do it again?

I mean ... come on!

@krishnanrohit Speaking from an account with ZDR enabled: you don’t

They're setting up trusted-access programs for cyber + bio. Basically, a safeguarded model for everyone (Fable 5) – downgraded to Opus 4.8 if classifiers fire – the full-capability version for vetted users (Mythos 5).

They've also clearly put work into vision, where Claude has historically been poor. Fable 5 rebuilt a web app's source code from screenshots alone, and beat Pokémon FireRed with a minimal vision-only harness.

@krishnanrohit oh shit ur right

@krishnanrohit OAI has had the same policy since forever

@krishnanrohit @zacharynado i don’t think it’s that big of an issue. Many/most apps have such language to capture more data than they actually need. HIPAA has data deletion requirements, but the window is greater than 30 days

@krishnanrohit surprisingly, this will be a problem for the enterprises, and not anthropic

@krishnanrohit This also is very strange to change one sided. E.g. if this model comes into cowork it would effectively change the data sharing agreements no?

@krishnanrohit my guess is ~no, or at very least something custom

@krishnanrohit BAAs

@krishnanrohit they won’t for now. we’re not deploying it rn fwiw

@tenobrus @krishnanrohit I suspect this is actually similar to the curly effect in political science.