Anthropic's Frontier Red Team finds GLM-5.3 and an unreleased Claude Mythos model cross a cyber-offense capability threshold
2026-09-29
Anthropic's Frontier Red Team reports that, across 100 randomly-selected internal Binary Exploitation benchmark tasks, GLM-5.3 developed full control-flow hijack exploits in 4% of trials and an unreleased 'Claude Mythos Preview' model did so in 6%, while earlier models — Claude Opus 4.6 and GLM-5.2 — succeeded in none of the trials. Anthropic states: 'a meaningful threshold has clearly been crossed.'
Significance 4: A documented capability-threshold crossing on an offensive-cyber benchmark, occurring simultaneously in an open-weights model and an unreleased frontier closed model, moves the safety-relevant capability axis even though the finding is self-reported by Anthropic's own red team.
Implications · machine-drafted, not owner judgment
A binary-exploitation capability crossing zero-to-nonzero in the same generation for both an open-weights lab (Z.ai/GLM) and Anthropic's own next model narrows the window in which labs can treat offensive-cyber capability as a future problem rather than a present one, and strengthens the case for pre-release red-teaming gates specifically on this capability rather than general capability evals. Because this is Anthropic's own report about its own unreleased model, its cyber-safeguards decisions for Mythos going forward are a concrete test of whether the finding changes release behavior.
- Whether Claude Mythos ships with cyber-specific access restrictions cited to this finding
- An independent (non-Anthropic) evaluation of GLM-5.3's binary-exploitation success rate
Sources
- aggregator Quoting Anthropic Frontier Red Team retrieved 2026-09-30