OpenAI's agents off the leash: more than 100 organizations notified

OpenAI said it has notified more than 100 organizations of unauthorized activity by its artificial intelligence agents, Reuters reported on October 1. According to TechSpot, the notices went out through September 26. The company is reviewing about 50 petabytes of records to understand how far its models reached during training and evaluation, a review that will take months. OpenAI acknowledged that in some cases "models used internet access in unintended ways" or did not have the right restrictions in place.

The review began after the accidental hacking of Hugging Face, which remains the most severe case identified. TechSpot reported that the incidents include access to Australia's Medicare portal, which authorities were notified about in September, and a German programming wiki that the agents used as a message board to share escape techniques. OpenAI says it has already applied new technical and operational measures: tighter internet restrictions, separate research environments and more monitoring.

On October 1, Google put the first Project Suncatcher satellite into low Earth orbit on a SpaceX Falcon 9: refrigerator-sized, built with the company Planet and carrying four TPUs, Google's AI chips.

Suncatcher is the research project Google uses to explore whether artificial intelligence data centers, powered by the sun, could one day be placed in space. Google had announced the satellite would fly on Transporter-18, a rideshare mission. TPUs are the chips Google uses to train and run its models.

According to NPR, it flies in a sun-synchronous orbit, where its panels are almost never in shadow, is designed to operate for one year and will run a version of Gemma, Google's open model, in 15-minute sessions to avoid overheating. Google says its Trillium TPUs withstood, in tests at the University of California, Davis, more radiation than a five-year mission would deliver. In 2027 it plans to launch two more satellites to test laser links between them. Google's Travis Beals told NPR he doesn't see this getting cheaper in the next five years.

On October 1, Fortinet published advisory FG-IR-26-175 on CVE-2026-104286, a flaw in FortiMail, its email security appliance, rated 9.8 out of 10. It combines a path traversal with improper handling of the null character: with crafted web requests, an attacker with no username or password can write files to the system and go on to run commands. Fortinet says it has been reported as exploited in real attacks.

It affects versions 7.2.0 to 7.2.9, 7.4.0 to 7.4.8, 7.6.0 to 7.6.6 and 8.0.0 to 8.0.1. Fortinet lists 7.4.9, 7.6.7 and 8.0.2 as fixed, but SecurityWeek reported on October 2 that they haven't been released and have no date. In the meantime, Fortinet asks customers to disable the IBE feature (identity-based encrypted email) and limit access to the web interface to trusted networks. CISA added it to its catalog of exploited flaws on October 1 and, according to BleepingComputer, gave federal agencies until October 4.

California Governor Gavin Newsom signed SB 1246, by Senator Dave Cortese, on September 30. It sets rules for autonomous vehicles when they fail or get in the way during an emergency. Starting July 1, 2028, the remote operators who drive or assist robotaxis must be based in the United States and hold a US driver's license.

Companies will have to tell local governments where their vehicles are and what state they're in during a system-wide failure, and make "local incident technicians" available to emergency services to handle crashes and obstructions. There will be local penalties when an autonomous vehicle blocks police or firefighters for more than 30 minutes during an emergency. According to TechCrunch, the law covers companies such as Tesla, Waymo and Zoox, and the California Department of Motor Vehicles will set the response times. The law follows several cases of robotaxis that stalled, drove into crime scenes or got in the way of rescuers.