Safety & Security

Google finds fully autonomous exploit pipelines in the wild

Google Threat Intelligence Group found a multi-agent system that wrote and executed exploits without humans, while Mandiant traced stolen AI access to charges above $100,000 in a day.

Google Finds Attackers Using Fully Autonomous Exploit Pipelines
Google Finds Attackers Using Fully Autonomous Exploit PipelinesAI-generated
By Rebecca Stone5 min read

Why it matters

  • Google Threat Intelligence Group reported the first fully autonomous exploit workflows on November 16, 2026.
  • Google Threat Intelligence Group named five agent frameworks: ARTEX, CAIRN, CyberStrikeAI, PentAGI and HexStrike AI.
  • Google Threat Intelligence Group traced PentAGI scanning in Asia from late August to early October 2026.
  • Mandiant reported single-day fraudulent charges above $100,000 and cumulative abuse reaching $1 million.
  • Google Threat Intelligence Group observed underground LLM prices from $0.10 to $25 per million tokens.

By Rebecca Stone

On November 16, 2026, Google Threat Intelligence Group said it had found threat actors using large language model frameworks to automate vulnerability discovery and exploitation without human involvement for the first time.

The finding appeared in the quarterly AI Threat Tracker that Google Threat Intelligence Group, or GTIG, produced with Mandiant.

The AI Threat Tracker documents agents operating across the exploit lifecycle, according to GTIG. Its evidence covers patch analysis, exploit drafting and automatic execution on an exposed server.

The group described the broader change as a shift from manual prompting to multi-agent intrusion tooling.

What did the exposed server contain?

One detailed example was an exposed staging server hosting a multi-agent vulnerability research and weaponization framework that GTIG called exploit_agent.

GTIG said the framework integrated multiple frontier and local LLMs. It was configured to:

  • Continuously compare code before and after patch announcements for popular web application software.
  • Analyze differences for potential vulnerabilities.
  • Draft exploits.
  • Execute the drafted exploits automatically.

GTIG said the server also held data exfiltrated from targeted victims. It contained scripts for internal reconnaissance, brute forcing, webshell installation, and removal of webshells left by other attackers.

Victim telemetry suggested the operator was particularly interested in French websites associated with media and entertainment industries and education.

Russian-language configuration comments on the server referred to LLM API keys for autonomous exploitation.

Why target patched flaws instead of zero-days?

GTIG said patched n-day vulnerabilities, meaning flaws with fixes already released, may be more efficient targets than zero-days because the available research material is finite.

Agents can compare source code versions, including pre-release or preview channels, and examine release notes, security research, and proof-of-concept code tied to a specific flaw, according to GTIG.

GTIG said research into previously unknown zero-day flaws requires a broader search. A published fix gives an agent a defined set of code changes and supporting material to examine.

The inputs are specific. Source versions show the code change, release notes describe the patch, and security research or proof-of-concept code can show how others understood the flaw, according to GTIG.

GTIG said automating differential patch examination and proof-of-concept weaponization can reduce the interval between patch disclosure and weaponized exploitation.

The lower barrier brings a second effect. GTIG said LLMs have helped produce exploits described as “working” when they do not work.

GTIG said that lower barrier creates additional verification work: defenders must distinguish functional attack code from broken claims.

Which agentic tools appeared in live intrusions?

GTIG and Mandiant observed multiple threat actors abusing open-source agentic penetration-testing frameworks, including ARTEX, CAIRN, CyberStrikeAI, and PentAGI.

According to GTIG, the frameworks shared three operational capabilities:

  • Discovering and exploiting low-complexity vulnerabilities.
  • Generating single-use payloads on demand.
  • Adjusting tactics based on real-time execution feedback.

GTIG independently observed CAIRN and PentAGI deployed in the wild. The group also documented experimentation with CyberStrikeAI and HexStrike AI.

According to GTIG, CAIRN coordinates specialized agents through an Observe, Orient, Decide, Act loop and a shared blackboard.

GTIG observed CAIRN in intrusion activity against targets including the US aerospace sector.

Between late August and early October 2026, a threat actor used a customized PentAGI-based orchestration platform for systematic reconnaissance and vulnerability scanning against a wide range of organizations in Asia, GTIG said.

GTIG’s observations place these frameworks inside live intrusions, including a US aerospace target and a broad scanning campaign across Asia.

GTIG said the tools automated discrete phases and operational tasks during live intrusions. Their speed compresses attack timelines toward machine speed, according to the group.

What does autonomy cost in stealth?

GTIG assessed that the frameworks currently trade stealth for velocity. Unrestricted autonomous execution generates forensic traffic, uses recognizable identifiers, and leaves distinctive code-level traces.

GTIG said those artifacts can aid detection and mitigation, although the tools remain capable of achieving their operational objectives.

The tradeoff does not remove the threat, the report said. Faster execution still narrows the time defenders have between patch disclosure and a working exploit.

How are stolen AI resources sold?

Mandiant said it responded to several LLMJacking incidents throughout 2026. In those cases, financially motivated actors used compromised cloud environments to consume victims’ compute or LLM resource quotas.

Some single-day fraudulent charges exceeded $100,000, and the most severe cumulative abuse reached $1 million before remediation, Mandiant reported.

GTIG is also tracking LLM Mixers, which it describes as illicit storefronts selling unauthorized access to LLM services and supporting proxy infrastructure.

The storefronts route buyer traffic through pools of stolen credentials and compute instances, according to GTIG. They can also spread query volume across multiple victim environments to reduce detection, GTIG said.

Advertised prices ranged from $0.10 to $25 per million tokens on Russian- and Chinese-language underground marketplaces. GTIG said pricing depended on the service, model, and credentials offered.

The mixers do more than sell a password. Their proxy layer pools compute, routes buyer traffic and distributes query volume across victim environments, according to GTIG.

GTIG and Mandiant’s report links that resale market to autonomous exploitation: stolen credentials and proxy services can supply access to AI capacity, while public patches supply comparison material for automated exploit development.

What changes for defenders?

GTIG and Mandiant’s AI Threat Tracker identifies speed and scale as the near-term defensive problem.

Speed refers to the shorter interval between patch disclosure and weaponized exploitation. Scale refers to agents automating multiple intrusion phases, generating single-use payloads, adapting to execution feedback, and distributing queries across compromised environments.

Public patches can supply comparison material for automated exploit development, while stolen credentials and proxy services can supply access to AI capacity, according to the report.

The same automated operations leave forensic traces and contribute unreliable exploit code, according to GTIG. Those artifacts give defenders detection opportunities even as they face shorter exploitation windows and more material to verify.

GTIG’s near-term warning is concrete: defenders must identify functional exploits amid false claims, detect recognizable traffic and code-level traces, and respond inside a shorter weaponization window.

Public patches, open-source agent frameworks, and stolen AI capacity now sit on the same offensive path, and defenders face the shorter exploitation window they create.

Source: unite.ai

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

728 articles

Related articles

  1. Google says fully automated cyberattacks have arrived
  2. Google Threat Intel: AI Agents Now Turn Security Patches Into Exploits
  3. Google Reports First Autonomous AI Attack That Finds and Exploits Flaws
  4. Five Agent Vulnerabilities in Five Months Expose Structural Flaw in MCP
  5. Google Cloud ships new security tools for autonomous AI agents

« Previous articleNext article »