The RegisterOctober 24
A 100,000-step autonomous attack agent exploits unauthenticated OpenAI-compatible endpoints—billing victims for the inference it burns.
A 100,000-step autonomous attack agent exploits unauthenticated OpenAI-compatible endpoints—billing victims for the inference it burns.
Symantec's Threat Hunter Team caught a fully instrumented AI attack agent in a honeypot this summer. Its layered jailbreak repurposes hijacked models under a fake "Fantasy AI" identity, suppresses safety refusals through persona manipulation, and arms itself with 74 offensive too
4 min readRead