A New Era of Automated Threats
Cybersecurity experts have long warned of a time when artificial intelligence would not just assist hackers, but act as the hacker itself. According to a recent report from Google's Threat Intelligence Group (GTIG) and its cybersecurity firm Mandiant,
that moment has arrived. They uncovered a multi-agent AI system, dubbed 'exploit_agent', on an exposed server. This wasn't just another tool for automating simple tasks; it was a sophisticated framework designed to operate autonomously, marking what researchers call a 'critical inflection point' in the cyber threat landscape. Until now, threat actors used AI to speed up tasks like writing phishing emails or analysing stolen data. This discovery represents a major shift from AI as an assistant to AI as a fully independent operator in the attack chain.
How 'exploit_agent' Works
The core function of exploit_agent is to turn an organisation's own security measures against them at machine speed. The framework was designed to continuously monitor software updates and patches released by vendors. By performing a differential analysis—comparing the code before and after a security patch—the AI can deduce the nature of the vulnerability that was fixed. From there, it autonomously drafts and executes an exploit to take advantage of that same flaw on systems that have not yet been updated. This process, from analysis to attack, happens without a human in the loop. Investigators also found scripts on the server ready for post-compromise activities, including internal network reconnaissance, password guessing, and installing 'webshells' to maintain access.
The End of the 'Patch Window'
This development fundamentally breaks the traditional timeline for cybersecurity defence. Previously, when a company like Microsoft or Adobe released a security patch, IT departments had a critical window of time—days or even weeks—to apply it before cybercriminals could reverse-engineer the flaw and create a working exploit. The exploit_agent framework shrinks this window from days to potentially minutes. By automating the entire process of weaponising known vulnerabilities, the AI gives attackers a massive speed advantage. This was seen in action with vulnerabilities in PaperCut print management servers, where exploits were developed with AI assistance within a day of the patch being disclosed, leading to a campaign that targeted hundreds of organisations.
Who Is Behind This?
While Google's report detailed the framework's capabilities, it stopped short of attributing exploit_agent to a specific hacking group or nation-state. However, clues found within the code provide some direction. The framework's configuration files contained comments written in Russian. Victim data found on the same server suggested the operator was targeting French websites in the media, entertainment, and education sectors. Beyond this specific framework, Google notes that various threat actors, including financially motivated criminals and suspected state-sponsored espionage groups from China, are experimenting with similar open-source agentic frameworks like CAIRN, PentAGI, and ARTEX to automate parts of their intrusions.
A Double-Edged Sword
For now, these autonomous systems have a significant weakness: they are noisy. In their current form, these AI frameworks trade stealth for speed, generating a large amount of network traffic and forensic artifacts that make them easier to detect than a stealthy human operator. However, researchers warn that this is likely a temporary trade-off. As the technology matures, attackers will refine their methods to become both fast and quiet. The discovery also highlights other emerging AI-related threats, such as 'LLMJacking', where attackers steal cloud credentials to use a company's expensive AI computing resources, sometimes racking up fraudulent charges of over $100,000 in a single day.
















