Anthropic’s team, which had never tried it before, removed the safety refusals from the most cyber-capable open-weight model for about $4,400 of computing time.
That is from Anthropic’s own analysis of Z.ai’s GLM-5.3, published September 29. The model built a working exploit against Chrome’s V8 engine in 50 of 410 attempts, against 56 of 410 for Anthropic’s restricted Mythos Preview. A US government standards lab had already called it the most cyber-capable open-weight model released to date. The refusal-removal numbers come from Anthropic alone, in a simulated setting, so they still need an independent replication.
The same week, the White House published a safety accord that names no penalty and no enforcer. NVIDIA shipped a watchdog chip that quarantines a misbehaving agent from outside the machine it runs on. Both work for organizations that choose to adopt them. Neither does anything about someone running a downloaded model on their own hardware.
The pattern across this week’s issue is that anything made of software was copied, repriced, or promised within days. Cloudflare copied a new category of model in sixteen days. Silicon, electricity, and signed contracts kept their terms.
If your security plan assumes attackers are a generation behind the tools your defenders can buy, what is that assumption based on?
Full issue: https://www.teamignite.vc/blog/sixteen-days-to-copy-ten-years-to-pay