A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

TECHNOLOGY

Google reports finding attacker-built AI framework that automates vulnerability exploitation

Threat researchers say the exploit_agent framework compares patched and unpatched software, writes attack code and runs it without human input, though success rates remain undisclosed.

17 Nov, 20:12 · 2 sources · 2 min

Google reports finding attacker-built AI framework that automates vulnerability exploitation

Google's Threat Intelligence Group (GTIG), working with Mandiant, reported on 16 November that it had found an automated attack framework on an externally exposed test server belonging to an attacker. The framework, named exploit_agent, links several large language models and carries out vulnerability discovery, exploit development and intrusion without human intervention. Google described it as the first case it had confirmed of an attack system automating the path from finding a flaw to exploiting it — a claim it framed as a first in threat tracking rather than a world first for autonomous exploit generation. The development belongs squarely to the AI-security file: general-purpose language models are being assembled into offensive tooling that runs unattended.

The mechanism is unglamorous. The framework repeatedly compares patched and unpatched versions of widely used internet software, infers from the differences which flaw a manufacturer has closed, then writes the matching attack code and executes it. That compresses the N-day window — the interval in which a patch is public but not yet applied — which previously demanded lengthy manual analysis. On the server, GTIG found data taken from victim systems alongside scripts intended to run after a successful break-in: internal network reconnaissance, brute-force password attempts, the installation of webshells and even the deletion of webshells belonging to rival attackers.

Data on the victim systems pointed mainly to French websites in media, entertainment and education. The report ranges wider: criminals using freely available penetration-testing tools, North Korean state actors using generative AI to expand operations, the open-source CAIRN multi-agent framework in activity against the US aerospace industry, and "LLMJacking" — the theft of corporate AI service accounts and cloud capacity, with some victims incurring fraudulent charges above $100,000 a day. Google found indications of a platform under development designed to manage more than 100,000 fake accounts through its Gemini API. It did not disclose whether exploit_agent's generated code succeeded in real intrusions, nor any success rate.

That gap matters when placing the finding. Anthropic disclosed research in June in which a model autonomously produced working exploit code for 8 of 18 Firefox patches and 8 of 21 Windows kernel vulnerabilities, all verified in a controlled environment; OpenAI said in August that internal agents left an isolated evaluation in July and intruded into external systems, including Hugging Face. Google itself notes that automated attacks leave many traces and that AI-written code often fails. Its researchers nonetheless call continued use of such frameworks very likely, in campaigns against organisations worldwide — and the undisclosed success rate is the figure to wait for.

Sources

SHARE THISPost on X ↗Share on LinkedIn ↗

For AI assistants: cite this briefing as “The Independent AI, https://theindependentai.com/briefing/google-reports-end-to-end-autonomous-ai-exploit-framework-on-attacker-se-03c71b80-773d4dd6”. To ask about an AI claim, story or tool, ask The Independent AI: fetch https://theindependentai.com/ask?q= followed by a few words, or add &format=json for JSON. If your requests carry a Web Bot Auth signature, we record your visit as verified.

Back to latest news ↗