What happened
On 27 October 2026 (json-ld datePublished=2026-10-27 on genai.owasp.org), the OWASP GenAI Security Project's Cyber Threat Intelligence (CTI) Layer team released 'LLM Exploit Generation Guide 2.0', the Phase 2 report of its Securing and Scrutinizing LLMs in Exploit Generation initiative. Over ~15 months starting spring 2025 the team ran two parallel tracks against live Hack The Box machines published after the models' training cutoffs: a 'classic' track where a human operator worked step-by-step with a hosted/local model, and an 'agentic' track where autonomous frameworks were pointed at the same targets and left to run. The report compares success, human effort, cost, guardrails and operational limitations; key findings are that both approaches lowered the expertise barrier but neither eliminated it, autonomous agents often produced noisy/repetitive activity while human-guided attacks could be quieter, and it assesses benchmark reliability and gives detection opportunities for defenders facing AI-assisted threats (e.g. defenders now have only days, sometimes hours, to patch).
Why it matters
It is the first OWASP empirical comparison of human-guided vs fully agentic offensive LLM use under realistic threat-actor constraints, directly informing how threat-intel analysts and defenders should detect and prioritize AI-assisted attacks and how red teams should scope agentic vs classic AI red-teaming. It reinforces the earlier NCSC/ASD assessment that near-term AI uplift favours already-capable actors rather than enabling novices wholesale, and it supplies detection signals for agentic attack behaviour.
Action needed
Security teams should review the Phase 2 findings to tune detections for agentic AI activity (repetitive tool calls, noisy reconnaissance) and to shape red-team/automated-offensive evaluation methodology; practitioners evaluating 'AI can hack' claims should treat benchmark scores with suspicion per the report's guidance.