Researchers at Google Threat Intelligence Group said they have witnessed a threat actor using an AI agent to identify vulnerabilities and develop exploits without human involvement.
The researchers identified an LLM framework called exploit_agent which was configured with Russian-language tooling to continuously compare pre- and post-patch code, identify vulnerabilities, and automatically develop exploits without human intervention.
The observed example of no-human-in-the-loop crosses a troubling threshold yet seen in the sphere of autonomous cyber threats.
Researchers found that threat actors are increasingly leveraging LLMs and agentic frameworks to fully automate the reverse engineering of patches via differential analysis, allowing them to rapidly weaponise n-day vulnerabilities and scale exploitation pipelines against internet-facing servers without human intervention.
The Google Threat Intelligence Group observed financially motivated and likely cyber espionage threat actors using open-source agentic pentesting tools, including CAIRN, PentAGI and CyberStrikeAI, to automate steps across an attack.
The Group observed CAIRN used during intrusion activity against targets, including in the aerospace sector in the U.S.
Researchers found that state-backed groups and commercial “Influence-as-a-Service” vendors are using LLMs to build automated pipelines, design infrastructure, and evade detection.
Recommended reading
- Scottish Business Confidence Rebounds Amid Tech Investment Push
- Microsoft Reveals How Quickly AI Is Weaponising Vulnerabilities
- Investing in Women Code (2026) Report | Key Takeaways
- Ana Stewart’s Pathways Pledge Launches Seventh Cohort
In one case, actors used the Gemini API to build platforms designed to manage more than 100,000 synthetic accounts through cloud phone farms, including features that mimic smartphone camera flaws to make generated content look more authentic.
The Group observed China-aligned actors using LLMs to process stolen data into intelligence products.
It also tracked North-Korea-aligned activity targeting cryptocurrency firms with AI-generated personas, including some carrying SynthID watermarks, alongside efforts to manipulate the configuration of autonomous coding assistants.





