AWS Security Agent (AWS Continuum) exploits 33 of 37 freshly disclosed real-world CVEs end-to-end on a contamination-reduced benchmark

On Oct 15 AWS published results from its autonomously operating pen-testing agent (now part of AWS Continuum): against 37 live containerized targets from 2025-26 CVEs (no PoC/report given, no runtime CVE lookup), it captured 33 flags end-to-end via the intended vulnerability (89.2%), with 2 reclassified and 2 misses caused by destructive guardrails — plus a new internally built, continuously refreshed freshly-disclosed-CVE benchmark that disables both memorization and retrieval contamination routes.

This is a credible, independently audited measurement of an autonomous offensive AI at real-world exploitation, on fresh (non-memorized) CVEs spanning 8 vulnerability classes — stronger signal than saturated public benchmarks. It validates that frontier-agent employment of freshly disclosed vulns is close to production-real, and both buyers of agentic pen-testing and defenders of AI/ML platforms (themselves on the target list) should take note.

Security teams evaluating autonomous penetration testing / vulnerability validation agents for head-to-head benchmarks; AI-platform and web-framework maintainers should assume fresh-CVE exploitation by agents is feasible and harden accordingly.