Elektrine lite

← Feed

@emory@soc.kvet.ch

Post #1604062

2026-04-22 18:03 UTC

I don't really have a handle on this "models distributed on huggingface that intentionally write insecure code if your politics don't align with those of the trainer" situation, but it really erodes the argument that using local LLMs is better for privacy. #infosec #AI #cybersecurity #privacy #sleeperAgentic

Replies (1)

  • @emory@soc.kvet.ch they’re not so bad if folks would more commonly publish open-weights, but even that would only address poisoning problems and does not address the “most code is not secure enough therefore models produce insecure code” problem

    Open ##2366601