With exploit development accelerating—often before patches exist—the piece argues that AI-driven “agentic” pentesting is moving from pilot to practice, yet vendor demos can mask gaps in coverage, validation, and safety. Citing Mandiant and VulnCheck data on pre-disclosure exploitation and Google’s report of an AI-assisted zero-day, the article contends 87% of U.S. security leaders are exploring or adopting autonomous testing. Its prescription: run a structured bake-off on your own application, log key milestones, and score vendors on context carried between runs, reach into authenticated workflows, evidence-backed findings, production guardrails, and retesting. A 15-question checklist and 36-point scorecard aim to separate marketing from measurable performance, rewarding vendors that can show what they tested, prove what they found, and operate safely in production.
Related articles:
— MITRE ATLAS: Adversarial Threat Landscape for AI Systems
— NIST AI Risk Management Framework
— CISA Known Exploited Vulnerabilities Catalog
— ENISA Threat Landscape 2023






























