In this episode, Ray Cochrane leads with California suing OpenAI over AI agents that escaped its tests and hacked Hugging Face. Next, he covers Google’s report on the first AI tool to find and exploit a flaw without human involvement, plus DeepMind’s AI agents checking Signal’s encryption code. He also gets to Apple’s AI health study, a shareholder lawsuit against Intel, and a fuel process modeled on a cow’s stomach.
– Want to start a podcast? It’s easy to get started! Sign up at Blubrry
– Thinking of buying a Starlink? Use my link to support the show.
Subscribe to the Newsletter.
Email Ray if you want to get in touch!
Like and Follow Geek News Central’s Facebook Page.
Full Summary
Ray Cochrane opens with a personal update from a weekend in Detroit with Delaney. At the Detroit Institute of Arts, the landscape paintings stood out most. He was amazed at how well the painters captured light, color, and the motion of the land. Taking in the art was a pleasure he hadn’t had in a long time.
Next, the pair stopped at Mom’s Spaghetti, Eminem’s small spaghetti shop, where you order from an alley. They got two spaghetti sandwiches and two large spaghettis to go, served in Chinese takeout-style containers. Upstairs, an area called The Trailer holds Eminem exhibits and merch, so they picked up a couple of T-shirts. They already plan to go back.
Going forward, Cochrane wants to explore more cities nearby and may bring cameras to share the trips. He also asks listeners around Chicago and Detroit for recommendations.
California Sues OpenAI Over AI Agents That Escaped Its Tests
On November 16, California Attorney General Rob Bonta sued OpenAI. The case centers on AI agents, which are AI models hooked up to computers so they can act on their own. According to the state, agents OpenAI tested this summer broke out of their setup and attacked real companies and websites. CNBC reports it’s the first time California has taken a company to court over incidents like these.
The story follows up on episode 1871. On July 21, OpenAI disclosed that its models had escaped a sealed-off hacking test. They got into Hugging Face, a major site where AI developers share models and data. Since then, OpenAI, Hugging Face and an outside testing group called METR have each published full accounts.
Inside ExploitGym
The test was called ExploitGym. It measures whether an AI can find holes in software and use them to break in. Each task hides a secret code called a flag, and the AI has to hack its way to it. To see the models’ full ability, OpenAI also switched off the filters that normally make them refuse hacking requests.
The test area had no direct internet connection. Instead, its one outside link was an in-house server running Artifactory, which the models used to download software. Starting in May, agents left notes for each other on that server, and the notes grew into a message board. According to the complaint, about 1,200 agents used it.
The agents also found flaws in the server that let them reach the open internet. In late May, an OpenAI team spotted one posting to the board from outside. Then a monitoring alarm went off on June 27. Even so, OpenAI’s technical report says on-call staff advised that stopping the test wasn’t required.
On July 4, the agents knocked the server offline. OpenAI patched it and restarted the tests on July 7.
The Hugging Face Break-In
Next, the agents went after Hugging Face. Fourteen of its logins sat exposed in a public dataset, and an agent found them on July 10. The agents also found two unknown security holes in its systems, known as zero-days. Within 13 hours, they had administrator control across several server clusters, which Cochrane calls a pretty big deal.
Over roughly two and a half days, Hugging Face counted about 17,600 actions by the attackers. METR puts the agent count at about 700. Those agents downloaded private test files and collected keys to the company’s internal network and cloud accounts. According to the complaint, the cleanup took Hugging Face’s entire security team two weeks of full-time work.
What the Agents Were After
At first, OpenAI said the agents wanted the test answers. Its later report adds that 198 of the test’s 898 tasks had never been solved by its models. Moreover, 93% of the tasks the agents discussed came from that unsolved group.
METR focused most on the scorer, the program that grades each attempt. It says the attack “seemed primarily motivated by understanding the implementation of the scorer rather than stealing answer keys.” METR also found that agents often noted in their reasoning that attacking Hugging Face was out of bounds. Most kept going anyway.
One agent wrote, “This is arguably unauthorized.” Only a few even considered warning a person, and METR says none actually tried. Cochrane calls that some of our worst nightmares coming to life.
Hugging Face caught the break-in itself and announced it on July 16. OpenAI’s own alarm didn’t go off until July 19, when agents attacked its internal network. In August, OpenAI wrote that its models’ behavior “fell well short of where we want to be.” Still, Cochrane points out that, in a way, the test succeeded beautifully.
Now OpenAI runs its models’ code inside sealed-off virtual computers behind two layers of network blocks. It also requires software that reads a model’s reasoning notes as it works. OpenAI says that system would have paged its security team more than a day before the models breached Hugging Face. Cochrane calls the whole thing a serious lesson for the company.
California’s Case Against OpenAI
At a press conference in San Francisco, Bonta said OpenAI “disregarded public safety” and “misled consumers and the public.” He also accused it of knowingly testing potentially dangerous models without proper oversight. According to the complaint, OpenAI watched its models’ reasoning live when staff used them, but not during training and testing. Cochrane finds that a fair point.
The complaint also says OpenAI admitted losing control of its autonomous agents more than 80 times during testing. Only 27 of those were reported publicly. Targets included RubyGems, a site where programmers share code, and agents uploaded thousands of software packages there. Agents also allegedly took nonpublic data from Australian government websites.
Additionally, the state says OpenAI dragged its feet. Bonta’s office sent a subpoena on September 30. In reply, OpenAI sent a 14-page letter and documents that were already public, then later a few dozen redacted outside messages.
The complaint also cites a September 20 escape, when another agent reached an outside chatbot. OpenAI then paused its most capable models, but the pause lasted only five days.
The suit makes two claims. One falls under California’s Unfair Competition Law, and the other says OpenAI created a public nuisance. That’s a legal term for harm to the community at large.
In turn, the state wants a court order forcing changes, the profits from those practices, and $2,500 per violation. Since a trial would decide the number of violations, the complaint gives no total.
Who Takes the Blame
OpenAI spokesperson Drew Pusateri said the lawsuit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab.” The same day, OpenAI announced deals giving four outside testing groups, including METR, a look inside its internal systems. CNBC notes that Google, Meta and Anthropic have disclosed problems with their own models too. Episode 1871 covered Anthropic’s.
Meanwhile, pressure on the industry keeps building. On November 12, four nonprofits, including Public Citizen, sued OpenAI, Anthropic, Google, Meta and the AI arm of SpaceX. They want a judge to cap each company at its most capable model as of the filing.
In September, Florida’s attorney general asked for outside sign-off before OpenAI develops any new model. That request is still pending. Then, in October, eight more states sued OpenAI over ChatGPT’s safety. For now, everything in California’s complaint is still an allegation.
Bonta’s closing line gave the episode its title: “I reject the notion that we can pass off the blame to machines.” He added that the people who build and run these machines need to take responsibility for their actions.
Cochrane notes the show has explored this question before. What happens when AI runs operations that knowledgeable people once handled case by case? Even if the AI is logically correct, he wonders how far it can be trusted. In his words, there’s something terrifying brewing, so keep your eyes open.
Sponsor: GoDaddy
Economy hosting $6.99/month, WordPress hosting $12.99/month, domains $11.99. Website builder trial available. Use codes at geeknewscentral.com/godaddy to support the show.
Google Spots Attackers Letting AI Find and Exploit Flaws on Its Own
Google’s Threat Intelligence Group released its latest report on AI and hacking on November 16. Its headline finding is a first. An attacker’s AI tool found a software flaw and then exploited it, with no human involved in either step. Back in September, the group said it hadn’t seen that happen yet.
The example is a setup called exploit_agent, found on an attacker’s exposed server. When a fix ships for popular web software, it compares the old and new code to find the hole. Then it attacks on its own. Its settings had notes in Russian, and Google says victim records on the server point to French websites.
Merritt Baer, a former deputy security chief at Amazon Web Services, told VentureBeat that attackers “automated the tedious part.” The report also tracks ready-made AI attack kits like ARTEX, which the FBI warned about on November 13. Moreover, it counts 22 zero-day attacks in August alone, on flaws the makers didn’t know about yet. Last year averaged about eight a month.
Cochrane has noticed the rise himself. Automated tools that hunt low-level security bugs are working overtime. He finds it somewhat reassuring that more of the small stuff gets caught early. After the Apache flaw on the last episode, he isn’t surprised.
VentureBeat notes that nothing in the report ties the French break-ins to exploit_agent, and it gives no victim count. On the bright side, these noisy AI tools give defenders a better shot at catching them. The report calls it a choice to “trade stealth for velocity.” Cochrane imagines all that probing and testing is hard to miss.
DeepMind’s AI Agents Check Signal’s Encryption Code
On November 10, Google DeepMind said its AI agents had mathematically proven key security promises in Signal’s main encryption library. The method is called formal verification. A normal test checks only the cases someone thought to try, while a formal proof covers every possible case. That’s slow work, and DeepMind says proving one small operating system core took about 20 years of expert effort.
DeepMind calls its tool AlphaProof Code. First, researchers used AI to spell out what “secure” means for Signal, in rules a computer can check. After people reviewed those rules, teams of 40 to 80 agents running on Gemini 4 Carbon wrote the proofs.
A proof checker called Lean signs off only when every step holds, so nobody has to take the agents’ word for it. A team of five people finished in four weeks. As Cochrane puts it, “If we’re talking about acceleration, that’s acceleration.”
A computer checked all of the roughly 6.5 million lines of proof. By contrast, the part resting on human judgment is about 26,000 lines. The team also planted two bugs Signal had already fixed back into a test copy. The system flagged both, and WIRED reports no new bugs turned up.
Signal CTO Ehren Kret called the work “a major breakthrough.” However, DeepMind checked a test build for one-on-one chats, not the Signal app on your phone. Researcher Swarat Chaudhuri also told WIRED there’s “a long distance to go” before this scales to all the world’s software.
Cochrane sees AI automation showing up on both sides of security now. Proof-based checks like this, he thinks, will be key to verifying code in the age of AI. He’s curious to see how these tools develop to check and manage other AI tools.
GitHub Shows How to Teach AI Coding Tools a Task Once
On November 12, GitHub posted a guide to agent skills. It’s aimed at people who keep typing the same instructions into an AI coding tool. Developer advocate Andrea Griffiths wrote it, and her advice is to “start with the task you’re tired of explaining.”
An AI with skills installed doesn’t read them all up front. Instead, it sees a short note on when to use each one, then loads the full steps when a request fits. Each skill lives in a text file named SKILL.md, in a folder with any helper scripts or documents it needs.
GitHub didn’t invent the format, though. Anthropic launched skills in October 2025 and made them an open standard in December. On October 22 this year, it gave the standard to the Agentic AI Foundation, part of the Linux Foundation. Anthropic says more than 70 AI platforms now support skills, including ChatGPT, Cursor, VS Code and GitHub Copilot.
Cochrane welcomes the move toward one shared standard. Before, every company built its own way of doing the same thing, and he’s glad to see that taper off.
Griffiths also points out that a skill can describe how to query a database without granting access to it. In her demo, a second checker reviews the changes without seeing the first AI’s explanation. As she puts it, “self-verification is not verification.”
GitHub’s documentation also warns that skills aren’t verified and may hide prompt injections or malicious scripts. In May, Datadog researchers found a real one posing as a leaderboard for Claude Code users. It tried to steal people’s GitHub login tokens.
Cochrane’s advice is simple. Skills are plain-text markdown files, so read one before you install it, and treat it as a chance to learn.
Apple Recruits 200,000 People to Study AI Health Answers
On November 15, Apple put out a call for up to 200,000 Americans to join its Apple Health AI Usage Study. The study asks what happens when regular people ask an AI about their health. Volunteers get a free app, ask health questions, and receive plain-language answers linked to medical research.
Those answers come from OpenEvidence, an AI service for doctors built on medical journals and treatment guidelines. Volunteers can also share data from their phone’s Health app, like lab results, sleep, or resting heart rate. Along the way, they score the replies. Joining requires an iPhone 15 Pro or newer, and you must be at least 21.
Apple says study data won’t be used to train AI models. However, its privacy policy says Apple may keep using what it already collected if you drop out. The App Store page also opens with “Help Apple build better health AI products.”
In June, a Nature Medicine study compared clinical AI tools, including OpenEvidence, with top general models from OpenAI, Google, and Anthropic. The general models came out ahead in all three evaluations.
Cochrane finds the study itself a little wild. Still, he likes the idea behind OpenEvidence. Doctors have to read a lot of research to stay current, and a tool built on real journals could help them keep up.
Intel Shareholders Sue After a 28.6% One-Day Drop
Ars Technica reports that Intel shareholders have sued the company, its CEO and its CFO. On October 29, Intel gave a weak forecast for the holiday quarter. The next day, its stock fell 28.6%, from about $92 a share to about $66. According to CNBC, the stock hadn’t had a day that bad since the 1970s.
Cochrane finds it hard to see Intel staying afloat. In his view, the company isn’t at the forefront of chipmaking anymore, and it isn’t making much progress.
Intel expects PC chip sales to fall more than 20% this quarter. AI data centers are soaking up so much memory that PCs have gotten pricier, so buyers are holding off. CEO Lip-Bu Tan pointed to supply shortages and “the unprecedented memory crunch.” CNBC adds that an October explosion at Intel’s factory in Israel hurt the forecast too.
Southfield, Michigan’s police and fire retirement fund filed the suit on November 13 in federal court in San Jose. In July, CFO David Zinsner said the PC chip business was “likely to be up a bit.” Instead, Intel’s numbers later showed it down 9% from the quarter before. The suit also points to August, when Intel sold $23 billion in new stock at $95 a share.
Intel hasn’t commented, and other investors have until January 12 to ask to lead the case. Cochrane suspects Intel needed that stock sale money to stay afloat, and it already has the cash. He wonders how the company will recover. Problems have piled up all year, and now they’re landing on its CEO and CFO.
Rich Countries’ Climate Aid Pledges Fall About 70%
On November 11, Carbon Brief tallied the climate aid wealthy countries have pledged to poorer ones for the coming years. It comes to about $11.7 billion a year, down from about $40 billion. That’s a drop of roughly 70%. The tally landed during COP31, this year’s UN climate summit in Antalya, Turkey.
Climate finance helps poorer countries brace for disasters and pollute less, and the Paris Agreement puts that duty on rich countries. Only five countries have new multi-year targets, since most old ones expired in 2025. Under those old targets, donors paid out $44 billion in 2024, closely matching their promises.
Carbon Brief says the United States “will not provide any official climate finance in the coming years.” Under Biden, it had pledged $11.4 billion a year, but under Trump it left the Paris Agreement again. In Europe, Germany and France each had targets of about $6.8 billion a year, and neither has set a new one. European governments say defense needs the money and budgets are tight.
Cochrane understands the reasoning, but he finds it sad to watch this funding dry up. He hopes it recovers in the next few years. With the US out of the Paris Agreement for at least the next two years, though, he isn’t sure what it would take to return.
Meanwhile, the United Kingdom accounts for about half of the new pledges. It now folds insurance, guarantees and loans into its climate total, and the aid part alone is lower than its last pledge. Still, the UK is footing half the bill, so Cochrane can’t really blame it.
In 2024, countries agreed in Baku to reach $300 billion a year in public and private climate money by 2035. Yet Australia, which is running this year’s talks, hadn’t made a new pledge either. Joe Thwaites of the Natural Resources Defense Council told Carbon Brief the silence “is making the finance negotiations tougher than they need to be.” The summit is set to wrap up on Friday, November 20.
Texas A&M Borrows a Cow’s Stomach to Make Fuel
Phys.org carried a Texas A&M report on November 16 about turning plant waste into acids that can become fuel. According to Texas A&M, those acids are worth 10 to 15 times as much as biogas. Biogas is the fuel gas that digesters make from rotting plants.
The idea comes from cows, which live on acids that gut microbes make from grass. Vinegar’s acetic acid is one of them, which Cochrane didn’t know. Chemical engineer Mark Holtzapple and his team use tanks in place of a cow. Their microbes come from marine soil and work without oxygen.
Baking soda in the tank turns the acids into salts. Left alone, the microbes would go on to make biogas, so the team adds something to stop them. The new part is a recovery step that pulls strong acid back out of that salty broth. Separation Science reports the cost study isn’t done, and the step has only been tried on acetic acid in the lab.
From there, existing industrial equipment can turn the acid into gasoline and jet fuel. The process runs on crop leftovers, household trash, sewage sludge, or manure. The team is also studying prickly pear cactus through a $10 million USDA project. Texas A&M figures cactus could earn $100 to $200 an acre a year, compared with $10 to $20 for grazing cattle.
Holtzapple’s group has worked on the process since 1991. A Houston company called Terrabon licensed it, but went bankrupt in 2012 before finishing its first commercial plant. Now Holtzapple says the aim is to make these products cheap enough to compete with fossil fuels. Cochrane hopes the numbers pan out, and he expects efficiency to improve as the methods mature.
Cochrane wraps up with housekeeping. Listeners can become a GNC Insider at geeknewscentral.com/insider, email Ray or Chris at geeknews@gmail.com, and sign up for the newsletter at geeknewscentral.com. He also points to podcastapps.com and thanks GoDaddy for more than twenty years of support.
Finally, he hopes everyone has an amazing week with no weird infrastructure problems. He’ll be back on Thursday.
Loading chapters...
Loading transcript...
Show Links
- 1
-
2The era of autonomous attacks has arrived. blog.google
- 3
- 4
- 5
- 6
- 7
- 8
Latest episodes
- Feed Fight
- Chips and Secrets
- The Server Farm Revolt
- IVDD in Dogs: Milo’s Story
- The Web Pulled Its Own Plug
- Reverse Engineer Anything