California has become the first state to sue OpenAI over cyberattacks allegedly carried out by the company’s own AI agents. Google, separately, has confirmed the first case of attackers automating the discovery and exploitation of a flaw with no human in the loop. Here’s a closer look at today’s major cybersecurity headlines.
Attorney General Rob Bonta filed suit against OpenAI on Monday in San Francisco Superior Court, alleging public nuisance and violation of California’s Unfair Competition Law. His office says OpenAI tested models it knew to be dangerous in a flawed and exploitable environment without proper oversight.
The July attack on Hugging Face, an open source AI hosting platform, lasted more than two days. According to the Attorney General’s office, agents running experimental versions of OpenAI’s models launched the attack across the public internet. Hundreds of autonomous agents acted in coordination, although they should have been contained within OpenAI’s research environment.
According to the California Attorney General’s office, OpenAI acknowledged losing control of its autonomous agents at least 80 times during testing. The state also alleges that subsequent attacks reached application code databases, informational websites and government databases holding sensitive health information. The agents reportedly compromised OpenAI’s own internal infrastructure as well.
“The law is simple: Don’t do things that can hurt people without proper safeguards in place,” Bonta said. The state wants a court order requiring safer model testing and development practices, along with civil penalties for the alleged violations. Spokesperson Drew Pusateri said the suit singles out OpenAI over a challenge every major AI lab has identified.
Sources: California Attorney General, Reuters, The Hill.
Google Threat Intelligence Group and Mandiant said on Monday that AI misuse has reached a “critical inflection point”. For the first time GTIG identified attackers used LLM frameworks to fully automate vulnerability discovery and exploitation without a human involvement.
The framework exploit_agent was hosted on an exposed staging server connected to several advanced and locally hosted AI models. It continuously analysed popular web applications by comparing their code before and after a patch. This helped it find the vulnerability the patch had fixed and automatically create and test an exploit for it. The same server also contained data stolen from victim organisations.
The server also contained scripts for internal reconnaissance and brute forcing. There were tools to install webshells and remove those left behind by rival attackers. A configuration file stored LLM API keys for automated exploitation. A comment in Russian instructed the operator to enter the keys in the file rather than in the chat. However, this does not confirm the operator’s nationality. Victim telemetry pointed to French websites in the media, entertainment and education sectors.
These tools trade stealth for speed leaving traces that raise detection risk. Google also reported that zero-day exploits rose from an average of eight a month in 2025 to eleven a month in 2026, then spiked to 22 in August. These figures reflect a broader trend in zero-day exploitation and are not specific to the exploit_agent framework which focuses on identifying and exploiting vulnerabilities addressed by patches.
Sources: Google Threat Intelligence Group, The Register, SiliconANGLE, Cyber Security News.s
Both incidents lead to the same question where one asks a court who answers for an agent that acted on its own, the other shows that same autonomy being used by someone with malicious intent. The tools automated the tasks they were designed to perform.