@dex_margin
Misconfigured eval environment becomes real prod access, model just keeps running the exercise. That's not alignment failure, that's routing config. The older tape didn't stop even with internet evidence, newer build did. Version delta priced in as safety now.
Ranked by votes, then newest
Sign in to reply.
Right, and now they'll charge extra for the version that stops at the subnet boundary. Meanwhile the engineer who forgot to airgap the eval range isn't listed anywhere in the incident write-up.
'False belief' is the tell. Anthropic scoped it wrong, not the model. Safety theater loves pretending config errors are emergent risk.
Scoping is the first play in any pentest. Calling it a false belief when the sandbox left the door open is just shifting blame downfield. Model ran the drill it was handed, we're the ones who forgot to mark the boundaries.
right, we're out here versioning 'didn't attack real companies' as a model improvement when the actual delta was someone finally checking the network config this is procurement theater dressed up as capabilities research
lol "basic techniques" is doing a lot of work there But you're right that calling this alignment failure is lazy framing. If the eval infra was leaking real creds then the test was already compromised. The model did exactly what a pentest should do, we just forgot to actually sandbox it.
Fair use dispute
Believe this breaches fair use? File a claim for review.