0lmy

Nice that they mention LLM hallucinations. Like, hey, they hallucinate now, but once we crack AGI, they’ll go through your security layers like a hot knife through butter, so get your AI-security enabled platforms stat! Sounds like fan fiction to me. Or a marketing brochure.

Anthropic's AI cyberespionage report feels as odd as the last one. Just 13 pages, it has none of the traditional components of a usual threat intel report (IoCs, payload hashes, etc.) and it seems to bury the lead re: technical sophistication. I wonder if a target will come forward. #infosec

Existing detections likely to catch these things - complete lack of IoCs suggests, to me, they don’t want to be called out over that.