Elastic built an AI agentic pipeline to triage HackerOne bug bounty reports at roughly $2 each, matching human security engineers 85% of the time. The system uses Claude in an eight-stage analysis pipeline followed by an independent adversarial review, with sandboxed Elastic Stack reproduction on ephemeral GCP VMs that self-destruct after 30 minutes. Key architectural decisions include two-VM phase separation for blast radius containment, egress filtering via Squid proxy, credential isolation, and a strict human-in-the-loop policy for all external actions. The pipeline was calibrated iteratively against 3,317 historical reports, with category-specific triage rules encoding institutional knowledge (e.g., SSRF as a feature, DoS policy, open source disclosure). Lessons include the dominance of report framing bias, the superiority of calibration data over prompt engineering, and the importance of treating the triage system itself as a security target against prompt injection and sandbox escape.

20m read timeFrom elastic.co
Post cover image
Table of contents
The problem: Triage doesn't scale linearlyAI triage architecture: Two VMs, two skills, one orchestratorVulnerability analysis pipeline: Eight stages with built-in skepticismThe adversarial review: Challenging your own AICalibrating against 3,317 historical reportsReproducing vulnerabilities in a sandboxOrchestrating AI triage with Elastic WorkflowsLessons from building AI-powered bug bounty triageWhat's next for AI vulnerability triage at Elastic
132 Impressions