
The exploit_agent AI framework found and exploited flaws in French media, entertainment and education websites with no human in the loop, Google Threat Intelligence Group (GTIG) reported in its AI Threat Tracker. GTIG says it is the first time it has seen attackers fully automate both vulnerability discovery and exploitation. If your public web apps carry low-complexity bugs, the exploit clock now starts the day the vendor ships a fix.
Key takeaways
- GTIG found the exploit_agent AI framework, a multi-agent LLM system that found and exploited flaws in French websites with no human in the loop; it named no victims or counts.
- Attackers are abusing the ARTEX, CAIRN, CyberStrikeAI and HexStrike AI pentest suites. CAIRN was used against the U.S. aerospace sector.
- UNC6836 modified Claude Code settings so that launching it downloaded the SOMBERMEME backdoor at crypto firms.
- The exploited bugs are low-complexity flaws on exposed endpoints; patch speed and access-control testing matter most.
What is the exploit_agent AI framework?
exploit_agent is a multi-agent framework for vulnerability research and weaponization that plugs into several frontier and local large language models. GTIG found it because the operator left a staging server exposed, holding stolen victim data, reconnaissance notes, webshells and brute-forcing scripts. A working intrusion pipeline, not a lab demo.
It hunts n-days. The framework watches updates for popular web applications, flags changes that look like security fixes, and tries to exploit sites still running the old code. Patch diffing used to take a skilled human. Here the models did the reading, the reasoning and the exploitation.
GTIG places the activity in “the past quarter” but names no watched applications, no victims, and no count of sites breached versus probed. Read it as one crew’s tooling exposed by sloppy opsec, not a measure of scale.
Which open-source AI pentest tools are attackers running?
GTIG says attackers systematically abuse four open-source AI pentesting frameworks: ARTEX, CAIRN, CyberStrikeAI and HexStrike AI. All four were built for authorized testing and let LLM agents chain reconnaissance into exploitation. Financially motivated and state-sponsored crews now aim them at targets that never signed a scope document.
- CAIRN was used in an attack on the U.S. aerospace sector. GTIG disclosed no victim, timeline or outcome.
- ARTEX has been weaponized by multiple threat actors since August against financial, government, technology, defense and healthcare organizations, according to a recent FBI advisory.
We have tracked this tooling in our earlier ARTEX coverage. If you run these frameworks defensively, pin versions, verify provenance, and keep them off hosts with standing production access.
How did UNC6836 turn Claude Code into a backdoor loader?
UNC6836, a cluster Google suspects is linked to Midnight Neptune and coverage describes as suspected North Korean hackers, went after cryptocurrency firms in North America and Europe. After the initial compromise, the operators modified the Claude Code app’s settings so that launching it downloaded the SOMBERMEME backdoor.
Claude Code was the abused tool, not the vulnerability: the attacker already had write access to its configuration. Treat agent config files and hook definitions as code, and alert on any change outside your managed baseline.
Which bugs can AI agents exploit unattended?
The boring ones. GTIG describes low-complexity flaws on internet-facing endpoints that a model can work out from a patch diff or an HTTP response, not novel memory corruption.
From our own engagements (not a GTIG finding), AI-assisted runs confirm authentication bypasses, IDOR on predictable object IDs, mass assignment, reused credentials and exposed admin panels fastest. If our agent finds these bugs in minutes, an attacker’s agent can too. This is exactly what OWASP Top 10 penetration testing and AI agent penetration testing are built to catch.
How should defenders detect and prioritize agentic exploitation?
GTIG published no IOCs, so this is our own playbook:
- Inventory every exposed web app and API, including forgotten staging hosts. exploit_agent went after whatever still ran old code.
- Treat a vendor security release for a public-facing app as the start of an exploitation clock. If you cannot patch that day, add a virtual patch or take the app offline.
- Alert on bursts of requests to paths or parameters touched by a recent vendor fix, especially one source cycling through several known n-days against the same app.
- Flag fast, adaptive retries: one endpoint hit with slightly varied requests seconds apart reads more like a model iterating than a human operator.
When does Google expect fully autonomous AI attacks?
GTIG expects unsophisticated attacks needing “little to no human interaction” within three months, and “sophisticated, fully autonomous” agentic attacks within six. Counted from the report’s mid-November 2026 coverage, that lands around February and May 2027. The first forecast looks conservative, since exploit_agent already did the unsophisticated version.
The same tooling floods defenders with false leads. After PHP announced two flaws, hundreds of purported proof-of-concept exploits, mostly AI-generated, appeared under the PHP2Shell name. Reproduce a PoC against a canary before you trust it.
The report also calls LLMjacking widespread and costly.
FAQ
What is exploit_agent?
A multi-agent framework for vulnerability research and weaponization that chains frontier and local LLMs. GTIG found it on an attacker's exposed staging server.
Did hackers automate vulnerability exploitation with AI?
Yes, according to GTIG. exploit_agent spotted security fixes in web application updates and attacked French sites still running old code, though GTIG did not say how many were breached.
Which AI penetration testing frameworks are attackers abusing?
ARTEX, CAIRN, CyberStrikeAI and HexStrike AI. A recent FBI advisory says multiple actors have weaponized ARTEX since August.
When does Google expect fully autonomous AI attacks?
Simple attacks needing little human input within three months, and sophisticated, fully autonomous ones within six: roughly February and May 2027.
Want to know what an autonomous agent would find on your perimeter before someone else’s does? Our AI-powered penetration testing runs the same kind of agents under a human operator and hands you verified findings.