AI guardrails block security research; devs switch to open GLM
OpenAI and Anthropic now refuse exploit and payload work; Hugging Face itself reached for GLM-5.2 to survive an attack its US model wouldn't examine.
Copy markdown
The hard stop on 'security-related'
OpenAI and Anthropic's frontier models now bail the instant a task reads as offensive security — writing exploit code, weaponizing a bug, even routine vuln assessment. A researcher at a phone maker put it flatly: "If it catches wind we're doing anything security related, it just stops and isn't usable."
Vetted access, or you're locked out
The official path is now gated: OpenAI's Trusted Access for Cyber and Anthropic's Cyber Verification Program, which CrowdFense's CTO says treat paying customers "like children." Anthropic's Mythos and Fable also caught June export controls over jailbreak fears, tightening the squeeze further.
Hugging Face's mid-incident refusal
Hit by a fully autonomous attack — tens of thousands of actions, 17,000+ log entries — Hugging Face's US frontier model refused to examine the malicious payloads, so the team fell back to Z.ai's GLM-5.2 to map the breach. CEO Clem Delangue: you can't have your tools refusing to look at payloads during a live incident.
Where the researchers are going
The workaround is Chinese open weights run locally — GLM and peers, no vetting, no refusals. RemoteThreat's Chris Thompson warns the guardrails are pushing "responsible researchers away from U.S.-governed systems to foreign-owned systems." If your work touches security, keep a local open model in the stack.