AI risk watchWarnings from authoritative sources
Anthropic warns that an open-weight model can now build working exploits with its safeguards stripped
Anthropic's Frontier Red Team found that GLM-5.3, an open-weight model from China's Zhipu AI, builds end-to-end exploits nearly as often as Claude Mythos Preview (50 of 410 attempts on ExploitBench, against 56), and that removing its refusals from the public weights took about $4,400 of computing without significantly reducing what it can do. Anthropic says the release "is a meaningful step change in the cyber capabilities available to attackers"; NIST's Center for AI Standards and Innovation calls it the most cyber-capable open-weight model released so far, about four months behind US frontier models.
Why it matters
Safeguards on a hosted model do not travel with downloadable weights, so the exploit-building skill is now available without anyone's permission.
What you can do
- Patch internet-facing systems first and fastest, above all those with public exploit details.
- Assume a published proof of concept will be turned into a working attack within days.