Researchers at Google and Mandiant have documented the first known instances of threat actors leveraging large language model frameworks to fully automate the discovery and exploitation of vulnerabilities. According to findings released Monday, attackers are deploying orchestrated multi-agent systems that coordinate specialized AI components to identify weaknesses, craft exploit code, and attempt execution against targets.

The automated exploitation workflow

Google's investigation uncovered a publicly accessible staging server hosting a multi-agent framework called exploit_agent, which contained stolen victim information and post-intrusion scripts. The system integrated multiple frontier and locally deployed language models to automate the comparison of patched and unpatched code, converting potential security gaps into functional exploit code ready for deployment.

The same server held tools for reconnaissance activities, credential-guessing campaigns, webshell installation, and removal of competing attackers' artifacts. Researchers identified indicators suggesting the operators targeted French organizations in media, entertainment, and education sectors, with particular focus on access control mechanisms within the Spectra WordPress plugin.

What's changed is that agents can now do much of that work themselves.

Comparing code before and after security patches to identify exploitable flaws represents a longstanding attacker technique. The significant shift lies in the capacity of AI agents to execute this analysis independently and at scale. Google declined to specify which models performed particular tasks or provide metrics on exploit success rates. While the server contained compromised data, researchers could not establish whether exploit_agent was responsible for obtaining it.

Offensive frameworks built on agent architecture

Beyond exploit_agent, Google identified attackers deploying open-source offensive security frameworks structured around orchestration patterns derived from coding agent design.

CAIRN operates by coordinating specialized agents through a shared information store and an observe-orient-decide-act decision cycle, using output from each operation to guide subsequent actions. Google observed this framework in intrusions against U.S. aerospace entities.

Alternative frameworks employ distinct strategies. CyberStrikeAI uses a Go-based orchestrator to manage commercial and open-weight models across more than 100 security tools, YAML-formatted attack recipes, and proprietary scripts. HexStrike AI makes more than 150 offensive security tools available through MCP, with dedicated agents specializing in reconnaissance, exploitation, and binary analysis.

Patch disclosure outpacing defense deployment

Vulnerability information now reaches attackers faster than defenders can deploy protective measures, compressing the window available for remediation.

Following exploitation of two PaperCut vulnerabilities as zero-days in August, security researchers and Huntress produced functional exploits with LLM support within a single day of public disclosure.

The attacker moved from initial access to domain administrator within minutes.

GreyNoise tracked a campaign launched four days after initial disclosure where an attacker likely leveraged language models to create exploits, compile target inventories, and compromise hundreds of organizations. The transition from first network access to administrative control occurred in minutes.

Two PHP vulnerabilities, CVE-2026-109618 and CVE-2026-109619, created comparable urgency when patched source code and a proof-of-concept demonstration surfaced before the official October 21 release date.

Following patch deployment, numerous purported exploits circulated online, though many proved nonfunctional and bore hallmarks of LLM generation. Security teams faced the challenge of distinguishing genuine threats from inert code.

Speed prioritized over operational security

Current-generation agent frameworks remain relatively straightforward to identify. Google determined that these systems produce substantial volumes of network traffic and leave behind detectable signatures and code patterns. Researchers observed that attackers "currently trade stealth for velocity."

The agents frequently generate incorrect technical information and reference nonexistent targets, sometimes disrupting campaigns or revealing operator identities.

For application security teams, the primary risk centers on the compressed timeline between patch publication and functional exploit availability.