Google Finds First Fully Autonomous AI Exploit Pipeline
By @sharedot · · 8 pages
- Cybersecurity
- Agentic AI
- Threat Intelligence
- Exploits
Google's GTIG reports the first fully autonomous AI, exploit_agent, turning security patches into working exploits with no human involvement.
What happened
Google Threat Intelligence Group, working with Mandiant, published its quarterly AI Threat Tracker on November 16, 2026, titled "Agentic Threat Activity Crosses a Significant Threshold."
It documents the first known instance of threat actors using a fully autonomous AI framework to convert public security patches into working exploit code with no human in the loop.
The framework, dubbed exploit_agent, was discovered on an exposed staging server wired into multiple frontier and locally hosted large language models.
It continuously compared code in popular web applications before and after each patch announcement, analyzed the differences for likely flaws, drafted exploits, and executed them automatically. The same server also held data exfiltrated from targeted victims.
Why it is a first
Two months earlier, the previous edition of GTIG's tracker stated it had not observed fully autonomous attack pipelines deployed against real targets in the wild, according to coverage by SiliconANGLE and The Times of India. That changed over the past quarter.
The shift moves beyond AI-assisted coding to an end-to-end automated pipeline in which agents handle vulnerability analysis, trial-and-error debugging, and final attack generation.
GTIG notes attackers may find patched n-day flaws more efficient targets than zero-days because the source material — patch diffs, release notes and proof-of-concept code — is finite and high-probability, allowing exploitation at machine speed.
The evidence on the server
The exposed staging server contained more than the exploit framework itself. Alongside exfiltrated victim data, it held scripts for internal reconnaissance, brute-forcing, installing webshells, and removing webshells left by rival attackers.
Victim telemetry suggested the operator focused on French websites tied to the media, entertainment and education sectors. A Russian-language comment in the configuration file instructed the operator to enter LLM API keys "right here, not in the chat," per unite.ai and The Times of India.
How fast a patch becomes a weapon
The report points to PaperCut Software print management server flaws, chained and exploited as zero-days in late August: security researchers and Huntress Labs used LLMs to produce working exploits within a day of disclosure, per The Times of India.
A separate campaign tracked by GreyNoise Intelligence four days after disclosure is believed to have used LLM-built exploits and hit hundreds of organizations, with one foothold reaching domain administrator within minutes.
unite.ai also reports LLMs have produced exploits described as working when they are not, adding verification work for defenders.
Agentic tooling and stolen AI access
Beyond exploit_agent, GTIG and Mandiant observed actors abusing open-source agentic penetration-testing frameworks including ARTEX, CAIRN, CyberStrikeAI and PentAGI.
CAIRN, which runs an Observe-Orient-Decide-Act loop, was used in an intrusion against the US aerospace sector, while a customized PentAGI platform scanned organizations across Asia between late August and early October, unite.ai reports.
Mandiant also responded to LLMJacking incidents with single-day fraudulent charges exceeding $100,000 and cumulative abuse reaching $1 million, and tracked "LLM Mixers" storefronts selling stolen LLM access at $0.10 to $25 per million tokens.
The stakes and what comes next
The core defensive problem is speed: autonomous agents can drastically compress the traditional patch window organizations have to apply updates, per OODA Loop, demanding automated defense-in-depth and accelerated patch management.
There is a counterweight for now — GTIG assesses these frameworks trade stealth for velocity, leaving forensic traffic, recognizable identifiers and code-level traces defenders can detect.
Google said it disabled accounts and assets tied to Gemini misuse it found, and Google DeepMind is using the findings to strengthen its classifiers, according to SiliconANGLE.
Sources
- oodaloop.com › Google Threat Intelligence Group Identifies First Known Instance of Autonomous AI Framework Turning Security Patches Into Working Exploits
- unite.ai › Google Finds Attackers Using Fully Autonomous Exploit Pipelines
- timesofindia.indiatimes.com › Google threat team finds first AI hacking without human intervention
- siliconangle.com › Google says an AI framework turned security patches into exploits with no human in the loop