Google’s Threat Intelligence Group says attackers are using autonomous AI agents to reverse-engineer security patches, generate exploits, and attempt to execute them without human approval.
In a report published Monday, Google and Mandiant describe the first cases their researchers have identified of attackers using LLM frameworks to fully automate vulnerability discovery and exploitation.
The framework uses an orchestrator to coordinate specialized agents, share context, and invoke tools through interfaces such as MCP. The agents apply familiar coding workflows to offensive security, identifying vulnerabilities, generating exploit candidates, and testing them against targets.
Inside the exploit pipeline
Google found a publicly exposed staging server running a multi-agent framework called exploit_agent alongside stolen victim data and scripts used after a successful intrusion. The framework connected several frontier and locally hosted language models to automate patch diffing, turning potential vulnerabilities into exploit code that it could then attempt to execute.
The server also contained scripts for reconnaissance, brute-force attacks, installing webshells, and removing those left by competing attackers. Google found evidence that the operators were interested in French media, entertainment, and education websites, and that the framework had examined access controls in the Spectra WordPress plugin.
What’s changed is that agents can now do much of that work themselves.
Patch diffing is nothing new. Attackers have been comparing code before and after security patches for years to figure out what was fixed and how to exploit it. What’s changed is that agents can now do much of that work themselves.
Google didn’t identify which models handled individual tasks or disclose how often the generated exploits succeeded. And while the server contained stolen data, researchers couldn’t determine whether exploit_agent obtained it.
Agent harnesses go offensive
Google also found attackers using open-source offensive frameworks built around orchestration patterns familiar from coding agents.
CAIRN coordinates specialized agents through a shared blackboard and an observe-orient-decide-act loop, using the results of each task to determine what happens next. Google observed the framework in intrusions targeting U.S. aerospace organizations.
Other frameworks take different approaches. CyberStrikeAI uses a Go backend to orchestrate commercial and open-weight models across more than 100 security tools, YAML attack recipes, and custom scripts. HexStrike AI exposes more than 150 offensive security tools through MCP, with specialized agents handling reconnaissance, exploitation and binary analysis.
Faster than patch cycles
Publicly available vulnerability information is also reaching attackers faster, leaving defenders less time to deploy fixes.
After two PaperCut vulnerabilities were exploited as zero-days in August, security researchers and Huntress developed working exploits with LLM assistance within a day of disclosure.
The attacker moved from initial access to domain administrator within minutes.
GreyNoise documented a campaign four days after the initial disclosure in which an attacker likely used language models to develop exploits, build target lists, and attack hundreds of organizations. The attacker moved from initial access to domain administrator within minutes.
Two PHP vulnerabilities, CVE-2026-109618 and CVE-2026-109619, presented a similar risk when patched code and a proof-of-concept video became public ahead of the official October 21 release.
Hundreds of purported exploits appeared online after the patches shipped, although many didn’t work and were likely generated with LLM assistance. Security teams then had to determine which exploits posed a real threat and which were simply nonfunctional code.
Velocity over stealth
These frameworks are still relatively easy to detect. Google found that they generate substantial network traffic and leave behind recognizable signatures and code artifacts. Researchers noted that attackers “currently trade stealth for velocity.”
The agents also hallucinate technical details and nonexistent targets, sometimes derailing attacks or exposing operators.
For application security teams, the concern is how quickly attackers can move from a published patch to a working exploit.
The agents also hallucinate technical details and nonexistent targets, sometimes derailing attacks or exposing operators.
The post Security patches are supposed to protect you. Google found AI agents turning them into exploits. appeared first on The New Stack.