Z.ai's GLM-5.3 edges Anthropic's Mythos 5 on CyberGym, trails ExploitBench
Z.ai, a Chinese AI firm, just announced its open-source model GLM-5.3 as a strong contender against Anthropic's Mythos 5 in cybersecurity tests.
GLM-5.3 actually beat Mythos 5 on the CyberGym benchmark (84.5% vs. 83.8%), but fell behind on ExploitBench (54.4% vs. Mythos's 78.0%).
Z.ai to release GLM-5.3 with safeguards
GLM-5.3 is better at spotting software vulnerabilities, though it completed fewer timed tasks than Mythos 5 (130 vs. 247).
Z.ai plans to release GLM-5.3 publicly soon, but will keep its most sensitive features locked down for safety.
The company is pushing responsible AI use with its "Open Source Shield" initiative, adding code-auditing functions to its ZCode programming product and systems to screen risky requests, monitor the model's work, and train it to reject malicious tasks.