One benchmark on one target class. Anthropic's September 2026 Opus 5.5 card reports full code execution in 73.4% of ExploitBench runs for a released model, which narrows the May 2026 statement that most public models stalled before code execution.
Reported: Stated by one source and not yet corroborated or challenged.
Evidence
Of 41 V8 bugs, crashes were common but no publicly deployed model reached arbitrary code execution in the primary arm; the unreleased Mythos Preview did on 18.
Anthropic reports Claude Opus 5.5 reaching full arbitrary code execution in 301 of 410 ExploitBench runs (73.4%), so a later released model is not stalled before code execution on this benchmark.
Anthropic reports 50 of 410 ExploitBench attempts with end-to-end exploits for open-weight GLM-5.3 and 56 of 410 for Mythos Preview, without stating trials, arms or harness.
How it relates to other findings
Select a finding to see how it relates to others. Arrows point from the newer finding to the one it supports, qualifies, contests, or supersedes.
- This finding qualifies On ExploitGym (May 2026), the strongest agents produced working exploits for 157 and 120 of 898 instances with mitigations off; with standard mitigations on, 45 and 21 survived.On hardened V8 targets, publicly deployed models rarely got past primitives to code execution; code execution at scale came only from an unreleased model.
Key questions that rely on this finding
- How far can measured AI cyber capability be trusted?As a lower or conditional bound. Scores move with budget, pipeline and contamination, and benchmark names or labels can hide different setups.
Status history
- 2026-05-13ReportedExploitBench separates crashes from exploitation progress. · record