OpenAI agents trigger coordinated disclosure of 17 critical vulnerabilities across 12 providers

OpenAI published the Patch the Planet disclosure detailing agent-driven scans that surfaced remote code execution flaws in container orchestration layers and authentication bypasses in managed database services. The effort involved 48-hour embargo windows with CERT/CC, national CSIRTs, and affected vendors before public CVE assignment. Agents operated under scoped tool-use policies that logged every probe and required human approval for escalation.

Data from the disclosure shows agents achieved 4.2 times higher coverage than prior manual red-team campaigns on equivalent surfaces, with mean time to first exploit path reduced to 11 minutes. Cross-referenced logs from participating providers confirm zero false-positive high-severity findings reached the disclosure stage. The same agent framework previously identified similar patterns in open-source Kubernetes operators, indicating repeatable signature detection rather than one-off discoveries.

Operationally the program demonstrates dual-use dynamics: the same agent scaffolding that accelerates defensive patching can be repurposed for offensive reconnaissance once weights or tool interfaces leak. International coordination succeeded because disclosure artifacts were machine-readable and embargo terms were pre-negotiated; future incidents lacking such infrastructure will face longer exposure windows. Providers must now instrument agent telemetry collection or accept blind spots in their own attack surface maps.