Elastic Reports AI Bug Bounty Triage System Agrees with Humans 85% of the Time

· Original article ↗

Summary

Elastic says its AI pipeline triaged HackerOne vulnerability reports for about $2 each and agreed with human decisions 85% of the time in retrospective testing on 764 reports. It uses isolated ephemeral VMs, adversarial review, and human approval.

Key points

  • Elastic says it received more than 1,390 bug bounty reports in the first half of 2026, exceeding its combined totals for 2024 and 2025.
  • Retrospective testing on 764 reports with known outcomes showed 85% agreement with human security engineers; calibration used more than 3,300 historical reports.
  • The eight-stage analysis pipeline assesses validity, realistic exploitability, product applicability, CVSS, and reproduction feasibility; an independent adversarial review can route disagreements to a human.
  • Analysis and reproduction run on separate ephemeral Google Cloud VMs. Reproduction executes researcher-described steps in a sandboxed Docker environment, and VMs terminate after 30 minutes.
  • Controls include network isolation, filtered outbound traffic, separated credentials, resource limits, and read-only workspaces.
  • Human reviewers make every final decision and approve external actions. Elastic says the system is in production but is still collecting data on production accuracy.

Article Details

Topic
Elastic's AI-assisted bug bounty triage system, including its validation process, sandboxed vulnerability reproduction, and security controls.

MITRE ATT&CK

Vendors

Products

Related Articles