Claude Opus 5 tops vulnerability recall benchmark

Deep investigation persistence in security agents maximizes vulnerability discovery but introduces a significant triage overhead that must be balanced with precision filtering.

• –**Unmatched Persistence**: Opus 5 utilized at least 27 of 30 actions in 24.5% of runs (compared to <2.3% for other models), spending extensive budget exploring adjacent attack vectors like CSRF and target flaw verification.
• –**Top Recall & Tail Bug Discovery**: Reaching 81.3% pass@3 recall (26/32 CVEs), Opus 5 was the only model capable of discovering three specific CVEs in the benchmark suite.
• –**Precision & Noise Penalty**: Opus 5 generated 121 rejected candidate findings versus Sol's 53, resulting in lower precision and giving Sol the higher overall F1 score.
• –**Optimization Vector**: Future agent tuning should focus on teaching models to self-assess when prolonged digging yields diminishing returns rather than forcing artificial early termination.