Regulation

Google warns of AI agents automating exploit development from security patches

Researchers at Google's Threat Intelligence Group have documented attackers deploying autonomous AI agents to reverse-engineer patches and generate working exploits at unprecedented speed, collapsing the window defenders have to deploy fixes.

3 min read
Security patches are supposed to protect you. Google found AI agents turning them into exploits.

Attackers are leveraging autonomous AI agents to reverse-engineer security patches, create functional exploits, and deploy them without human intervention, according to research from Google and Mandiant released Monday. The findings represent the first documented cases of attackers using large language model frameworks to fully automate the vulnerability discovery and exploitation pipeline.

The automated exploitation workflow

Google researchers discovered a publicly accessible staging server hosting a multi-agent framework called exploit_agent, along with stolen victim information and post-intrusion scripts. The framework connected multiple frontier and self-hosted language models to automate patch diffing—the process of comparing code before and after security updates—and convert potential vulnerabilities into executable exploit code.

The server also held scripts for reconnaissance activities, brute-force attacks, webshell deployment, and removal of competing attackers' shells. Researchers found indicators that the operators targeted French media, entertainment, and education sectors, with the framework examining access controls in the Spectra WordPress plugin.

An orchestrator coordinates specialized agents within the framework, distributes context between them, and invokes tools via interfaces such as MCP. The agents replicate standard software development workflows but apply them to offensive security: pinpointing vulnerabilities, producing exploit candidates, and validating them against live targets.

Patch diffing itself is not novel—attackers have analyzed code changes for years to understand what was patched and how to circumvent it. However, "What's changed is that agents can now do much of that work themselves." Google did not specify which models performed individual tasks or provide metrics on exploit success rates. While the server contained stolen data, researchers could not confirm whether exploit_agent obtained it.

Offensive frameworks built on agent orchestration

Google identified attackers deploying open-source offensive frameworks constructed around orchestration patterns derived from coding agent architecture.

CAIRN manages specialized agents through a shared blackboard and an observe-orient-decide-act loop, using each task's output to guide subsequent actions. Google detected this framework in intrusions against U.S. aerospace organizations.

Other frameworks employ distinct strategies. CyberStrikeAI uses a Go-based backend to manage commercial and open-weight models across more than 100 security tools, YAML attack recipes, and custom scripts. HexStrike AI provides access to more than 150 offensive security tools through MCP, with specialized agents managing reconnaissance, exploitation, and binary analysis tasks.

Patch disclosure outpacing defense timelines

Vulnerability information is reaching attackers with increasing speed, compressing the timeframe available for defenders to implement protective measures.

When PaperCut vulnerabilities were exploited as zero-days in August, security researchers and Huntress produced working exploits with LLM support within a single day of public disclosure. GreyNoise documented an attack campaign four days after initial disclosure in which an attacker likely used language models to build exploits, generate target lists, and compromise hundreds of organizations. According to the documentation, "The attacker moved from initial access to domain administrator within minutes."

Two PHP vulnerabilities, CVE-2026-109618 and CVE-2026-109619, posed comparable risks when patched code and a proof-of-concept video surfaced before the official October 21 release date. Following patch deployment, hundreds of purported exploits circulated online, though many were nonfunctional and likely generated with LLM assistance. Security teams faced the challenge of distinguishing genuine threats from inert code.

Speed over operational security

These frameworks remain relatively straightforward to identify. Google observed that they produce substantial network traffic and leave behind detectable signatures and code patterns. Researchers noted that attackers "currently trade stealth for velocity."

The agents also generate hallucinated technical details and nonexistent targets, occasionally disrupting attacks or revealing operator identities.

For application security teams, the central challenge is the velocity at which attackers can transition from a published patch to a functional exploit. The compressed timeline between disclosure and weaponization fundamentally alters the risk calculus for vulnerability management.

Source: The New Stack · Reporting supplemented by The Silicon Ledger staff.