Anthropic: GLM-5.3 hits 4% cyber-exploit rate, a new threshold

The comparison the Frontier Red Team draws is not between GLM-5.3 and Claude Mythos Preview in isolation, but between this generation and the previous one. When Claude Opus 4.6 and GLM-5.2 score zero on the same 100-task sample, the appearance of any successful full control flow hijack marks a step change rather than a marginal gain. The team's word choice — a meaningful threshold — signals that it treats the existence of these successes, not their size, as the headline.

The benchmark itself is internal and the task sample is small and randomly drawn, so the 4% and 6% figures are best read as evidence of capability presence rather than as precise capability rankings. Anthropic's framing acknowledges this by noting GLM-5.3 performs below Claude Mythos Preview while still describing the overall result as a crossed threshold.

The stakes, from this body of evidence, hinge on whether the successes hold up across a wider sample of tasks. For those tracking AI and cyber risk, the thing to weigh is not which model leads but that the floor has risen from zero — and the Frontier Red Team has put its own name behind that reading.