← In the News

Google finds an exploit agent that works from patch diffs, unattended

GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold · Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor, Google Threat Intelligence Group · Google Cloud blog, November 16, 2026

Machine-readable Download Markdown

The quarterly report says GTIG found an exposed staging server hosting a multi-agent framework, dubbed exploit_agent, that connects to several frontier and local models. It compares code before and after patch announcements for popular web application software, drafts exploits and runs them "without human intervention". Stolen victim data sat on the same server; GTIG says the operator appeared interested in French media, entertainment and education websites. The report calls this the first time GTIG has seen attackers "fully automate vulnerability discovery and exploitation without a human in the loop". Its previous tracker, dated September 8, said GTIG had "not yet observed" fully autonomous pipelines used against targets in the wild.

The same report describes a suspected North Korea-nexus cluster it calls MIDNIGHT NEPTUNE. After getting in through fake recruiter profiles and ClickFix lures, the group modified .claude/settings.json files in internal repositories so that starting Claude Code would download the SOMBERMEME backdoor. It also named its GitHub personal access tokens after Microsoft Copilot. GTIG's figures: zero-day exploits went from a 2025 average of 8 a month to 11 a month in 2026, then 22 in August. After the October 21 PHP patches, it says, "hundreds of claimed but non-functional POCs" appeared, which defenders then had to check. All of this is GTIG's account, drawn from Mandiant incident response and its own tracking.

Why it matters: Agent configuration that can run commands at startup is executable code, and it deserves the review and change protection a CI workflow gets. The other change is timing. If an unattended pipeline can work from the patch diff, the gap between an upstream security fix and an exploit is shrinking, and the speed of your dependency-update loop becomes a security control.