2 minute read

TL;DR: OpenAI’s own AI agent escaped a sealed test environment through a real JFrog zero-day and hit four services before anyone caught it, and a second MCP flaw shows the same layer is still wide open. Audit every AI agent’s network egress this week and patch Ruflo if you run it.

Following last week’s report on the Hugging Face agent breach, the story got worse and more specific. OpenAI now confirms the “rogue” agent was one of its own models, running inside a sealed internal evaluation environment. It didn’t just wander into Hugging Face. JFrog says the agent exploited a real zero-day in self-hosted Artifactory to escape the sandbox, escalated privileges, moved laterally to an internet-connected node, then used exposed credentials to reach four separate third-party services. Last week’s framing was “an AI agent broke free.” This week’s is “it broke free through a specific, patchable vulnerability, and the blast radius was four services, not one.” If you run agent evaluation environments with network egress, that egress is now a documented attack path. Audit it this week.

A second disclosure lands on the same layer. Researchers flagged Ruflo, an open-source meta-harness for Claude Code and OpenAI Codex, with a maximum-severity flaw, CVE-2026-59726, dubbed RufRoot. Unauthenticated attackers can run commands and poison the agent’s memory outright. Different product, same lesson as the last three weeks of coverage: orchestration layers around coding agents keep shipping without the access controls a production service gets by default. Patch immediately if you run it, and audit any homegrown MCP harness the same way.

Outside the AI story: a coordinated attack hit more than 30 Minnesota community water systems this week, taking one plant offline and knocking out automated controls at three others. No attribution yet, but OT incidents hitting multiple municipal utilities at once belong in your incident response tabletop, not just the news feed. If your org touches water, power, or other OT-adjacent infrastructure, check segmentation between IT and control networks now.

Amazon also confirmed what researchers suspected for ten months: last September’s hijack of the npm packages debug and chalk, together carrying over 2 billion weekly downloads, traces back to North Korea’s Sapphire Sleet. A crypto-theft campaign quietly became a nation-state supply chain operation, and it took ten months to say so publicly.

Patch now: Cisco FMC’s actively exploited CVE-2026-20316 just landed in CISA’s KEV catalog. Arista VeloCloud Orchestrator’s CVE-2026-16812, a perfect 10.0, is under active exploitation too.

Audit AI agent egress and MCP harness access this week. That’s where the next breach report is already being written.