How we built Datadog’s agentic vulnerability hunting system
Datadog describes an orchestration pipeline that equips AI agents with structured security context to hunt for vulnerabilities at scale. The core innovation is ‘threat packs’—concise, human-reviewed documents containing security invariants, trust assumptions, and compensating controls—that prevent agents from missing or over-reporting findings that existing controls already mitigate. The pipeline routes detection, reproduction, and validation tasks to cost-effective models and validates findings against explicit security requirements rather than raw code patterns. This approach reduced noise and improved report quality, uncovering concrete guest-to-host escapes in Kata Containers (virtio-fs) and Dragonball (virtio-blk) that required understanding architectural boundaries and trust assumptions that code alone wouldn’t reveal.