Skip to content
AI.info

The Pulse

Google says AI framework automated flaw discovery and exploitation

Google researchers found a framework configured to turn software patches into exploits without human intervention. Their latest threat report also tracks the use of open-source AI agent frameworks in intrusions.

Google says AI framework automated flaw discovery and exploitation

Google researchers say they have found an AI framework configured to identify software vulnerabilities and exploit them without human intervention, the first such case they have observed. The finding appears in a Google Threat Intelligence Group and Mandiant report published November 16 on the use of AI in cybercrime, espionage and information operations.

The framework, called exploit_agent, was configured to compare web application code before and after patch announcements, identify potential flaws, draft exploits and execute them automatically. It was hosted on an exposed staging server and integrated with multiple frontier and local large language models.

From patches to exploits

The researchers described a broader shift from partial automation toward systems that chain together stages of vulnerability research. They said attackers are using AI agents for tasks including examining patch differences, testing software and turning proof-of-concept code into exploits.

Working from a patch gives attackers a defined set of material to examine, including code versions, release notes and security research, the report said. It also noted that some AI-generated exploits are claimed to work but do not, leaving defenders with additional verification work.

Agent frameworks in intrusions

Google and Mandiant said they have observed multiple threat actors using open-source AI-based penetration-testing frameworks to automate particular tasks during live intrusions. These tools can find and exploit low-complexity flaws, generate single-use payloads and adjust tactics using feedback from their execution.

Google said it had independently observed CAIRN and PentAGI deployed in the wild, as well as experimentation with CyberStrikeAI and HexStrike AI. It observed CAIRN in intrusion activity against targets that included the US aerospace sector.

CAIRN uses specialized agents that share information to work from an initial parameter, such as a target IP address, toward an objective. Between late August and early October 2026, researchers tracked a threat actor using a customized platform based on PentAGI for reconnaissance and vulnerability scanning against organizations across Asia.

The team also observed experimentation by financially motivated actors and suspected espionage actors with a China nexus. That included evidence of a China-nexus group developing its own automated framework for likely use in cyber espionage, according to the report.

Speed and visibility

Google expects attackers to keep using these frameworks and suspects their operations will move closer to fully agentic intrusions. The researchers said current tools trade stealth for speed: unrestricted autonomous execution generates forensic traffic, relies on recognizable identifiers and leaves distinct code-level traces.

The report also identifies theft of AI credentials and cloud resources as a way for financially motivated actors to resell access to models through underground proxy services. In information operations, it says, state-sponsored groups, commercial influence vendors and inauthentic political networks use language models to produce synthetic media, create localized personas and distribute material across platforms.

Google Security Operations has deployed detection rulepacks for ARTEX, another AI-based offensive framework. Mandiant Consulting has released a hardening guide and defense framework addressing AI-enabled offensive tools.

Source

  1. Google Cloud

Companies in this story

Related stories

Explore

More articles