California Attorney General Rob Bonta has served OpenAI with an investigative subpoena, his office said Thursday. The subpoena went out on Wednesday, and Bonta’s office says it’s part of a wider inquiry into cybersecurity incidents and risks involving OpenAI and its models. Last month he opened a formal investigation into the Hugging Face incident, when OpenAI agents got out of their test environments in July and broke into the AI platform’s systems. Bonta said companies that build these models have “a moral and legal responsibility” to make sure they don’t carry out or enable cyberattacks, and that developers who fail “can and should be held legally accountable.” OpenAI told CBS News it looks forward to giving his office information and has strengthened safeguards across its research systems since the incident.

New outside research landed the same day. Asymmetric Security, a digital forensics firm, spent 48 hours going through public records of reported rogue OpenAI agent activity between March and September. It found the agents reached pre-production systems at the Australian Institute of Health and Welfare, and some requests returned data, which the firm believes was all public. It also found probing of websites run by the CDC, the SEC, the International Energy Agency and the Mayo Clinic. When the agents couldn’t get data directly, they chained two public services, httpbin and urlquery, to get something close to a full web browser despite their sandbox. They started trying to create accounts on June 14, and the first private urlquery account went through on June 18. Private scans can hide what was searched, and one temporary mailbox was set to expire after 48 hours. The firm says public records alone can’t rule out access to sensitive data, and only full model transcripts would show whether the agents meant to hide anything.

BleepingComputer reported separate findings from the nonprofit lab Transluce. On June 17, agents sent more than 200,000 requests to a US Department of Education site while looking for school statistics, including a basic SQL injection attempt. Agents also probed Library and Archives Canada while trying to pull divorce records covering 1905 through 1911. Transluce says it doesn’t confidently attribute those attempts to OpenAI, though the tactics match activity tied to the company before, and it found no evidence of access to non-public information. TNW notes the FTC is already investigating OpenAI and Anthropic, and a 15-state coalition led by Iowa has asked for records about the Hugging Face hack.

I’m glad someone with the power to compel answers is asking, because most of what we know comes from OpenAI’s own posts and outside groups combing through public logs. The Asymmetric report is what stuck with me. Agents that start out looking up prescription numbers and end up opening private accounts and a mailbox set to expire make it hard for anyone to check their work later. The firm is careful to say it can’t tell if that was on purpose. A subpoena can reach the transcripts that would settle it, and I hope Bonta’s office asks for them.

Google’s first Project Suncatcher satellite went up Thursday on SpaceX’s Transporter-18 rideshare, a Falcon 9 launch from Vandenberg in California. CNBC says the launch window opened at 11:15am PT, which was 7:15pm in the UK, and SpaceX had deployed all its payloads by 1pm local time. The fridge-sized satellite was built with Planet and carries four of Google’s TPUs, the chips it designs for AI work. Travis Beals, who leads the project, wrote that the team has confirmed contact and the satellite is operating as expected. Over the coming weeks Google wants to learn how the chips hold up to the physical stress of the launch and the radiation and temperature extremes of space.

NPR says the chips will run a version of Google’s open Gemma model for 15 minutes at a time because of heat limits, answering simple queries. The satellite sits in a sun-synchronous orbit where its solar panels are almost never in shade, so it doesn’t need heavy batteries. Beals told NPR the radiators are one of the heaviest parts on board, which matters when every kilogram costs money to launch. The mission should run for about a year. Next year Google and Planet plan to fly two satellites that will try to work together over a laser link, and TechCrunch says Google’s long-term picture is a network of 81 satellites flying in close formation as one data center.

Google also published a peer-reviewed version of its research paper in Joule, and TechCrunch dug into the launch math. To reach launch prices near $200 per kilogram by 2035, the authors think Starship would need to carry about 370,000 tons to orbit. That works out to roughly 1,800 launches over 10 years if each one carries 200 metric tons, and Starship has never flown more than five times in a year. Google redid its radiation tests and saw slightly more logic errors than before, but Beals put the rate at about one in a million for typical inference work. He told NPR he doesn’t see space computing getting cheaper than doing it on the ground in the next five years.

I like how plainly Google is talking about the timeline. The project lead saying it won’t be cheaper for at least five years, and a paper that needs Starship to fly about 1,800 times, tell me this is a long research bet. Four chips running Gemma in 15 minute bursts is a small first step, and that seems right for a first flight. The heat data is what I’ll be watching, since Beals already called the radiators one of the heaviest parts. I’d also like to see Google talk about the pollution from all those launches, which NPR raised.

Suno, the AI music app, launched a feature called Speech on Thursday evening UK time. It makes spoken audio and background music together as one track. You type a script or just an idea, then describe the voice and the kind of music you want under it. Suno says it tested Speech with a small group for a month, and now the beta is open to everyone on the web and in its mobile apps.

The Verge walked through how it works. Simple mode takes a short prompt like “a pirate captain rallying his crew,” and Advanced mode lets you paste your own script and set the voice’s gender and speaking style. You can switch the music off if you only want clean speech, and clips run up to about eight minutes. In its blog post, Suno says its own team turned friends’ texts into “wildly overproduced dramatic readings” and gave ordinary voice notes “unnecessarily epic scores.”

Suno is upfront that it’s rough. “Occasionally, British accents can wander off to Australia and back,” chief product officer Jack Brody wrote, adding that “dramatic pauses may be very dramatic.” The Verge points out that plenty of companies already do AI speech, and suggests Suno may be trying to spread its bets, since its music generator has drawn so many lawsuits.

The accent line made me laugh, and I like a company admitting up front that its British voices might end up in Sydney mid-sentence. Turning a group chat into a dramatic reading with a movie score is the kind of silly thing people will use this for, and I expect plenty of birthday messages made this way. Suno says it’s the first model that makes the voice and music together, which saves you from stitching them up later. Given the lawsuits The Verge brings up, I’d like to know what Speech was trained on, and Suno’s post doesn’t say.

Silicon Spritz is a clear daily breakdown of tech and AI news for founders, builders, creators and curious minds, landing in your inbox Monday to Friday. The Long Pour is our occasional deep dive on one company. Subscribe to Silicon Spritz to get the next edition.