AI · Security Desk · continually updated
AI security
The measured side of AI security: advertised bounty rewards, which vulnerabilities AI tools have found, how each frontier model does on security work and what it will answer, and the incidents as they are logged, with the reporting beside the data.
- programmes measured
- 1,143
- findings with public identifiers
- 100
- models scored
- 38
- incidents logged
- 266
- guides
- 5
The trackers
Each one downloads as CSV and JSON, CC BY 4.0Incident logAI Risk RadarThe standing log of AI-security incidents since 1 January 2026: prompt injection, hijacked agents, poisoned and leaked models and deepfake fraud, with the source on every record.266 incidents · 21 seen in the wildOpen AI Risk Radar Updated 28 Oct 2026Measured datasetBounty economicsWhat 1,143 public bug bounty programmes on seven platforms advertise as maximum rewards, their response times and safe-harbour terms, plus a selected comparison of AI-reporting policies.1,143 programmes · $3,625 median Bugcrowd ceilingOpen Bounty Economics Updated 28 Oct 2026Live registerAI-found vulnerabilitiesEvery vulnerability an AI tool found and a public record credits, each checked at its official CVE record or vendor advisory, with the strength of the evidence beside the credit.100 checked at a record · 124 entriesOpen the register Updated 28 Oct 2026Scored boardModel security capabilityHow each frontier model does on security work, with the harness, the task count and the safeguard state beside every score, and what each lab lets it answer.38 models · 65 scores · 18 independentOpen the board Updated 28 Oct 2026
