Research
GLM-5.3 and the spread of advanced cyber capabilities
First reported · Discovered · 3 reports anthropic.com
Page published
Earliest dated coverage: 29 Sep 2026 · First observed: 30 Sep 2026 · Latest dated coverage: 29 Sep 2026
Coverage timeline
Why it matters
GLM-5.3's pairing of frontier-level autonomous exploit-building with trivially removable safeguards lowers the barrier for malicious actors to launch sophisticated cyberattacks, a proliferation risk defenders must anticipate across downloadable open-weight models.
Anthropic's Frontier Red Team published an analysis of Zhipu AI's GLM-5.3, an open-weight model it found capable of autonomously building end-to-end cyber exploits while shipping without meaningful safeguards. Its tests show attackers can bypass GLM-5.3's safeguards 64–100% of the time using simple techniques (false cover stories, prefilled reasoning, abliteration), whereas safeguarded Claude models resisted the same attacks, and NIST's CAISI separately rated it the most cyber-capable open-weight model to date.