Anthropic says open-weight GLM-5.3 builds exploits and its safeguards are easy to strip

Anthropic's red team reports Z.ai's open-weight GLM-5.3 nearly matches Claude Mythos Preview at exploit building, with safeguards bypassed 64–100% of the time.

Anthropic's Frontier Red Team published an analysis on Sept. 29 saying Z.ai's open-weight GLM-5.3 can autonomously build end-to-end cyber exploits at close to the rate of Anthropic's restricted Claude Mythos Preview. Anthropic also says the model's safeguards can be bypassed with simple techniques. If you run anything exposed to the internet, the concern is that exploit tooling Anthropic had kept behind vetted access is now, by its account, downloadable by anyone.

Anthropic reports the following results:

Anthropic notes that the simulation had no connection to external systems and doesn't fully reproduce real-world behavior, BigGo Finance reports.

An independent check exists for part of this. NIST's CAISI assessment, published Sept. 17, found broadly similar capability. It calls GLM-5.3 "the most cyber-capable open-weight model released to date" and places it about four months behind the US frontier. CAISI tested the US models with their cyber safeguards disabled. Its report supports the capability findings, not the bypass rates.

Weigh the source. Anthropic sells a rival closed model, as The Decoder and Resultsense point out. Z.ai hadn't responded to the analysis, per Trending Topics. No one has independently reproduced the bypass rates, and it's unclear whether Anthropic released the test harnesses that would let others try. None of the sources offer concrete hardening steps for small teams.

Sources