exploit_agent framework
An attacker's AI pipeline, found by Google, that turned software patches into exploits and ran them with no human in the loop.
What happened
Google Threat Intelligence Group (GTIG) reported on Nov 16 2026 that it had found exploit_agent on an exposed staging server. The framework was wired to several frontier and locally hosted LLMs and continuously compared popular web applications’ code before and after each patch announcement; once it spotted a likely flaw, the exploit it drafted ran automatically. GTIG calls it the first fully autonomous attack pipeline it has seen deployed in the wild; two months earlier its tracker said it had not observed one.
The same server held stolen victim data and scripts for brute-forcing, installing webshells and removing rivals’ webshells. Victim telemetry pointed to French media, entertainment and education sites, and a Russian-language config comment told the operator to enter keys in the file, not the chat. GTIG argues attackers may find it more efficient to aim LLMs at n-day (already patched) flaws than to hunt zero-days.
The wider report
The quarterly tracker, “Agentic Threat Activity Crosses a Significant Threshold”, also describes:
- LLM-built exploits for the PaperCut zero-days within days of disclosure (see PaperCut AI agent campaign);
- the CAIRN multi-agent framework used in an intrusion against the US aerospace sector (see CL-STA-1238 Cairn campaign), and a customized PentAGI platform scanning organizations in Asia;
- a suspected North Korean group, UNC6836, whose fake recruiter photos carried Google SynthID watermarks and which altered
.claude/settings.jsonfiles so each launch of Claude Code downloaded a backdoor; - influence-operation platforms, one built with the Gemini API, designed to run 100,000+ social accounts.
Google says it disabled accounts tied to the Gemini misuse. It also notes that unrestricted autonomous execution leaves a noisy forensic trail that makes these frameworks easier to detect. See AI and CVEs: latent flaws, now exploitable for the broader pattern.
Timeline
Every dated event in the wiki that involves exploit_agent framework, newest first.
- Google threat intelligence reports the first fully autonomous AI exploit pipeline in the wild (exploit_agent)
Google Threat Intelligence Group's quarterly AI threat tracker describes exploit_agent, found on an exposed staging server: wired to frontier and local LLMs, it diffs popular web apps before and after patches and runs the exploits it drafts automatically, apparently against French media, entertainment and education sites. The report also cites LLM-built exploits for the PaperCut zero-days, use of the CAIRN framework against US aerospace, and a North Korea-suspected group (UNC6836) that rigged Claude Code settings files to fetch a backdoor.
Connections
exploit_agent framework and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
What links here
Concept AI and CVEs: latent flaws, now exploitable
Company Google
Incident PaperCut AI agent campaign