First no-human AI agent turned patches into exploits and stole data
Google's Threat Intelligence Group says a suspected Russian-speaking operator ran exploit_agent against common web software, with telemetry suggesting a focus on French media, entertainment and education sites.
Google's Threat Intelligence Group says a suspected Russian-speaking operator used an autonomous framework to turn security updates into working exploits without anyone guiding the process, and data from victim organizations was taken. Writing with Mandiant in its fourth-quarter 2026 tracker, the group called the case the first of that kind and a clear advance on what it saw in Q3.
The agent on an open server Investigators found a staging server exposed online. It ran exploit_agent, a multi-agent system for researching flaws and turning them into attacks, linked to several frontier models and to models hosted locally, with tools written in Russian. The system was set up to keep studying widely used web application software: it compared code from before a fix with the patched version, picked out weaknesses, built exploits and fired them on its own.
That server also held material taken from organizations that had been hit, plus several scripts meant for use after entry. Those scripts could map internal systems, guess passwords, plant webshells and remove webshells left by other intruders. Telemetry on the machine suggests the operator was focused on French sites in media and entertainment and in education.
Other crews and other schemes The same update says both criminals and espionage groups are leaving one-off prompts behind and fielding autonomous multi-agent attack kits. Named examples include the open-source testing tool ARTEX, plus CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. In a separate case, GTIG saw PentAGI paired with proxy services and aimed at aerospace. The group said it has watched several actors try these kits on their own, some of them after money and others suspected spies with a PRC nexus, and it cited how little it takes to start.
Elsewhere in the tracker, GTIG describes LLMJacking: intruders steal LLM logins and cloud capacity from breached companies, then resell use of high-end models through underground proxies it calls LLM Mixers. State-backed operators, vendors that sell influence work, and networks that fake politics for profit now regularly use LLMs to mass-produce synthetic media, build local-sounding personas and push posts across platforms. Espionage crews use the models for target research and for sorting collected data. GTIG also flagged a supply-chain approach in which attackers change how autonomous coding assistants are configured so a backdoor is dropped when code runs locally.