The Era of Autonomous Attacks Has Arrived
For years, experts have warned of a future where artificial intelligence could be used to launch cyberattacks. According to a new report from Google's Threat Intelligence Group (GTIG) and Mandiant, that future is now. Researchers have uncovered what they
call the first instance of a fully automated framework used by malicious actors for vulnerability discovery and exploitation. This marks a significant shift from AI simply being a productivity tool for hackers to AI acting as the operator, running attacks on its own.
How 'exploit_agent' Works
The framework, discovered on an exposed server and aptly named 'exploit_agent', connects to multiple large language models (LLMs). Its primary function is to perform continuous differential analysis. In simple terms, it automatically compares the code of popular web applications before and after a security patch is released. By analysing the changes, the AI can identify the exact weakness the patch was designed to fix. From there, it can autonomously write the code to exploit that vulnerability and launch an attack, all without a human in the loop. On the server, investigators found not only the tool itself but also data stolen from victims and scripts for post-compromise activities like brute-force attacks.
From N-Day to Zero-Day at Machine Speed
This new capability drastically shrinks the critical window between when a vulnerability is publicly disclosed (an 'n-day' vulnerability) and when it can be weaponized. What once took a skilled human researcher hours or days—sifting through code and crafting an exploit—can now be done in minutes by an AI agent that never sleeps. The report highlights how this allows attackers to scale their operations significantly, targeting systems that have not yet been updated. In one case involving vulnerabilities in PaperCut print management software, an attack campaign believed to be built with LLMs hit hundreds of organisations just four days after disclosure. In another instance, attackers gained top-level domain administrator privileges within minutes of their initial breach.
A Broader Shift in Cybercrime Tactics
The discovery of exploit_agent is part of a wider trend. Google reports that various threat actors, from financially motivated criminals to state-sponsored espionage groups, are adopting multi-agent AI frameworks. Tools with names like CAIRN and PentAGI have been used in attacks against targets including the U.S. aerospace sector and organisations across Asia. These agentic systems can automate different phases of an attack, from initial reconnaissance to executing complex intrusion chains. This shift means attacks are not only faster but also potentially more accessible to those with less technical expertise, as AI can handle the heavy lifting.
The Defensive Arms Race
The emergence of machine-speed attacks necessitates a machine-speed defense. Relying on human-speed responses is no longer feasible when facing an autonomous attacker. Google noted that just as attackers are using AI, defenders must embrace it as well. The company is developing its own AI-powered defense systems that use models like Gemini to continuously find and fix flaws before they can be exploited. The goal is to create an 'agentic defense' that can keep pace with the evolving threat. The challenge for organisations everywhere is to update their security posture, prioritising rapid patching and proactive threat hunting to stay ahead in this new, accelerated arms race.
















