GLM-5.3 and Claude Mythos show gains in exploit-generation tests

On September 30, 2026, Simon Willison published results from Anthropic's Frontier Red Team, which evaluated several models on 100 tasks selected at random from an internal Binary Exploitation benchmark. In that run, GLM-5.3 achieved full control flow hijacks in 4% of trials, while Claude Mythos Preview did so in 6%. Willison noted that although GLM-5.3 performs below Claude Mythos Preview here, a meaningful threshold has clearly been crossed: earlier models such as Claude Opus 4.6 and GLM-5.2 did not succeed on the benchmark at all.

Later the same day, the Chinese tech outlet OSChina reported on the same red team assessment with a different emphasis. According to its account, Anthropic's Frontier Red Team concluded that Zhipu's GLM-5.3 is the first model to be released with almost no refusal guardrails while being able to autonomously build end-to-end exploits in the way Claude Mythos Preview did five months earlier. The outlet recalled that when Anthropic released Claude Mythos Preview five months ago, it described that model as the first one capable of autonomously constructing complex end-to-end exploits.

The two reports agree that the red team evaluated GLM-5.3 and that Claude Mythos Preview had already cleared the same capability bar months before, but they frame the finding differently: the first centers on the 4% and 6% figures, the second on the absence of guardrails around the release. Neither account includes any response from Zhipu, and neither says what happens after the red team's assessment.