GLM-5.3 Crosses Frontier Cyber Threshold

The real story is not that GLM-5.3 has surpassed every U.S. model—it is that open weights have crossed a meaningful exploit-development threshold without equivalent safeguards.

• –Anthropic recorded successful exploits in 50 of 410 ExploitBench attempts for GLM-5.3, compared with 56 for Claude Mythos Preview.
• –GLM-5.3 reached 4% on Anthropic’s internal binary-exploitation benchmark, while earlier GLM-5.2 and Claude Opus 4.6 scored zero.
• –Anthropic found the model’s safeguards bypassable or removable, creating a substantially different risk profile from restricted U.S. frontier systems.
• –For security teams, the same capabilities could accelerate vulnerability discovery and patching—but only with strict sandboxing, access controls, and human oversight.
• –The competitive battleground is shifting from coding benchmarks to whether powerful open models can be deployed responsibly.