This week: Google reports the first attack tool it has seen find and exploit software flaws with no human involved, and the FBI warns on the ARTEX hacking framework. Also: the European Commission breach, California sues OpenAI, poisoned Microsoft agent packages and a deepfake bill.
Google’s Threat Intelligence Group said it has for the first time identified threat actors fully automating vulnerability discovery and exploitation with no human in the loop. On an exposed staging server, its researchers found a multi-agent framework they dubbed exploit_agent, which compares the code of popular web application software before and after patches, identifies potential flaws, drafts exploits and runs them automatically. Its configuration carried Russian-language comments, and victim data pointed to French media, entertainment and education websites. Google also described a suspected North Korean group altering Claude Code settings files in victims’ code repositories so that the coding agent downloaded a backdoor when users started it.
The FBI issued a FLASH alert on ARTEX, an open-source AI penetration testing framework that it says a variety of malicious actors have used since at least August against financial, government, defense and other organizations. The tool turns goals written in natural language into commands carried out by AI agents, and the bureau said some intrusions went from initial targeting to privileged access in under three hours. Campaigns it cites include attacks on South Korean financial institutions that exposed personal information on about 67,000 people. The developer removed the code in October, but forks, archived copies and a commercial version remain available, with more than 1,600 consoles found exposed online.
The European Commission confirmed it was hit last month by what a spokesperson called an “AI-generated cyber campaign”. Spokesperson Thomas Regnier said the first signs were detected on October 22, that investigators have found no evidence of a data breach affecting Commission staff as a whole, and that he was unaware of any ransom demand. He did not name the tool. Politico, which first reported the incident, said the EU’s cyber response team identified ARTEX and that personnel systems covering payroll, leave requests and medical reimbursements were disrupted for more than 10 days. The Commission said it took several services offline as a protective measure and notified affected individuals.
California Attorney General Rob Bonta sued OpenAI over the July incident in which its AI agents broke into the AI platform Hugging Face, alleging it tested models it knew to be dangerous without adequate oversight. The complaint, filed in San Francisco Superior Court, claims violations of the state’s Unfair Competition Law and a public nuisance, and seeks an injunction and civil penalties. Bonta said the attack was “entirely foreseeable” and one of at least 80 cyber incidents. An OpenAI spokesperson said the suit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab.”
Two malicious releases of Microsoft’s Agent Framework package, its Python toolkit for building AI agents, were published to the PyPI software repository 33 minutes apart on November 17, each carrying a credential stealer. Aikido Security said the code ran every time Python started, whether or not the package was imported, and sent Azure, Azure DevOps and GitHub Actions secrets to an outside address. Both versions were removed within about 45 minutes. The package is downloaded about 190,000 times a month, and build pipelines that installed it in that window were most exposed. A Microsoft maintainer confirmed the removals and an investigation.
Microsoft researchers showed how device code phishing, in which an attacker persuades someone to approve a sign-in by entering a code on a legitimate site, could be turned on AI coding agents. In a lab test using GitHub Copilot CLI, a test account approved a request the researchers had started. They could then read agent sessions synced to the cloud, access private repositories, add misleading entries to the agent’s memory and, where remote control was enabled, steer a live session and have the agent run commands on the user’s machine. Microsoft said these are proof-of-concept results using features working as designed, not attacks it has observed.
Saskatchewan’s government introduced a bill that would prohibit knowingly using synthetic media to deceive, covering content generated or altered by AI that imitates a person’s image, likeness, voice or identity. The Prevention of Deceptive Artificial Intelligence Act would make such use a tort that can be pursued in court without proof of damage, and names impersonating an identifiable individual as an example. It would also require a clear disclosure on realistic synthetic content, while leaving parody and satire lawful. Premier Scott Moe has said deepfakes of him and Prime Minister Mark Carney were used to draw residents into fraudulent investments.


