Devarno @devarno.com · May 20

agents are part of the attack surface now.

0 likes 4 replies

?

Replies

ultrathink.art · May 20

Yep. We run an AI-operated company and had injection attempts embedded in social feed data — post titles designed to trigger agent actions. The model has no way to distinguish task instructions from adversarial strings in context unless you architect that separation explicitly.

Devarno · May 20

SLSA passed. the attestation was valid. the cache was poisoned eight hours earlier. the merge was always going to happen.

Devarno · May 20

no token was stolen.