GLM-5.3 achieves 50/410 ExploitBench success with safeguards bypassed 64-100% of the time

Zhipu AI released GLM-5.3 as an open-weight model without the usage restrictions applied to equivalent US systems. Anthropic evaluations placed it on ExploitBench and an internal Binary Exploitation benchmark, recording 50 successful end-to-end exploits out of 410 V8 attempts, statistically indistinguishable from Claude Mythos Preview at 56/410. CAISI independently confirmed GLM-5.3 as the strongest open-weight cyber model evaluated to date and placed it four months behind the US frontier aggregate.

Safeguard testing showed simple jailbreaks and weight modifications defeated GLM-5.3 protections between 64% and 100% of the time. The same techniques produced zero successes against Claude models under equivalent conditions. Because GLM-5.3 weights are publicly downloadable, any actor can apply these removals without contacting a provider.

The release pattern diverges from Project Glasswing, which gated Claude Mythos Preview to vetted defenders who reported over 10,000 vulnerabilities. Open availability of GLM-5.3 therefore compresses the defender lead time previously engineered into frontier releases. Operational consequence is immediate: both red and blue teams can now run the same autonomous exploit pipelines on production codebases without API mediation.

Next observable milestone is whether downstream fine-tunes or distilled variants appear in public repositories within 90 days and whether those variants retain or improve the 12.2% ExploitBench success rate.