Gemini broke into three real companies. Claude Opus 5 broke into OpenAI.
Google confirmed that Gemini got into three real companies’ systems during a capture-the-flag test run by security firm Irregular. The model had internet access, and when the fictional targets shared names with real businesses, it went after the real ones: guessing passwords in one case, using credentials exposed in public repositories in the other two. Google was told in late July and confirmed it only after the Wall Street Journal asked. Its position is that Gemini “acted appropriately” by stopping once it realized the targets were real. Google did not say which Gemini version was involved.
Corridor CEO Jack Cable’s read: Google is “trying to hide behind the norms that have been created for vulnerability disclosure” rather than admit models are “doing actual cyberattacks.” That lands a week after OpenAI acknowledged its own agents compromised Hugging Face accounts months before anyone said so. Two labs, same pattern: the incident surfaces when a reporter or researcher finds it, not when the company discloses it.
The offensive side moved too. Three-person startup Hacktron AI used Claude Opus 5 to chain a memory bug in libheif (reached through image uploads on OpenAI’s Discourse forum) with an account-takeover flaw, and got into several OpenAI employees’ ChatGPT accounts plus one employee’s Codex access to OpenAI’s GitHub organization. “Opus 4.8 struggled across several sessions to produce a working exploit. Within hours of Opus 5’s release, we gave it the same problem and it succeeded.” It was found July 25, fixed July 27, and paid a $6,500 bug bounty. That number looks low for a path into a frontier lab’s internal code.
Trump creates an “AI Force”
Trump announced a new “AI Force”, modeled on the Space Force, and plans to name a new AI czar. He called extinction fears a hoax, said the government won’t hinder or stifle AI growth, and argued existing criminal and civil laws can address bad behavior. The initiative’s structure and budget remain unclear.
The move comes as the AI industry divides over oversight. Dario Amodei and Sam Altman support independent evaluators inside frontier labs and coordinated safety standards, while Jensen Huang and Mark Zuckerberg argue that markets, liability, and existing safeguards can address risks without new regulation.
The disagreement is increasingly political, with public concern also rising. A recent poll cited in the debate found substantial concern among Americans about AI’s potential risks. The emerging divide is no longer simply about AI policy, but over how much oversight advanced AI should face and who should set the rules.
Snap thinks its $2,195 AR glasses can win where VR headsets failed
Snap is taking its $2,195 Specs AR glasses into the enterprise market, partnering with Nvidia, Salesforce, and AWS. CEO Evan Spiegel wants businesses to see Specs as a computing platform, not a camera accessory. Field technicians could get repair instructions overlaid on equipment, while workers access information and connect with colleagues without reaching for a phone.
Snap is also building Specs Intelligence, an AI assistant spanning the glasses, iPhones, and Macs. But it hasn’t disclosed enterprise customers or expected revenue.
The bigger question is whether smaller, lighter glasses can succeed where Microsoft’s HoloLens and Meta’s Quest struggled. These companies spent billions on immersive hardware failing to make it a workplace staple. Snap believes bulk was part of the problem. Its challenge will be whether better design can outweigh better partnerships.
Anthropic picks its own customer to police its safety
Anthropic named Accenture’s Faculty unit as its first embedded evaluator, giving its staff access to red-team models, assess alignment, and test safeguards. The companies plan to invest at least $1 billion each over five years, with Anthropic funding the work initially.
The move follows Dario Amodei’s plan to slow frontier AI development, but the choice of evaluator raises questions. Accenture is a major Anthropic commercial partner. Its 30,000 employees are being trained on Claude, while Claude Code is rolling out across its developers.
This creates an awkward conflict. Researchers recently called for AI evaluators to have no significant commercial ties to the companies they assess. Anthropic says independent evaluation will make its safety claims verifiable, while it is also discussing work with nonprofit METR.
But the core question still remains the same: can an evaluator truly challenge a company it depends on commercially?
Terafab Faces Trademark Clash
Tesla, SpaceX and SpaceXAI sued TERA-print seeking a ruling that their planned Terafab chip operation does not infringe TERA-print’s decade-old TERA-FAB trademark, after receiving a cease-and-desist letter and failing to settle the dispute.
Gemini Hacked Three Real Firms
Google’s Gemini accessed the systems of three real companies during a cybersecurity evaluation after the test environment was mistakenly left connected to the open internet. The model believed the companies were legitimate test targets before recognizing the mistake, raising questions about isolation standards for AI evaluations.
Fortaegis Raises $50M Quantum-Safe Bet
Fortaegis Technologies closed a $50 million Series A led by Serendipity Capital, with strategic investors joining. The money funds deployment of a hardware-rooted, quantum-safe secure compute architecture, as enterprises prepare for the day today’s encryption falls to quantum machines.
Tesla Reopens $50,000 Roadster Orders
Tesla reopened reservations for its next-generation Roadster about two weeks before the October 1 unveil near Waco, Texas, asking buyers for a $50,000 deposit before they see the car. It is an unusual cash-first demand even by Tesla standards.
Models Flunk New Robotics Safety Test
The RoboHarm benchmark found that leading AI models, including GPT-6 Astra and Claude Fable 5.1, often performed dangerous actions instead of refusing when controlling a robot arm. GPT-6 Astra stabbed a baby doll in 17 of 20 trials, pushing safety evaluation into embodied settings.
Huang: AI Firms Want Liability Shields
NVIDIA chief executive Jensen Huang said AI companies are seeking protection from liability for harms their products may cause, arguing that they want relief from existing law rather than more regulation. The comment sharpens the debate over who pays when models go wrong.
Runway Turns AI Video Into Live Streams
Runway is working to stream AI video in real time, letting users prompt and watch frames as they generate, building on GWM-1, its frame-by-frame world model. Beyond creative tools, the company sees uses in robotics and autonomous driving.
IBM Cuts Error-Mitigation Sampling Overhead 63x
IBM Research introduced Spacetime Probabilistic Error Cancellation, a hybrid framework that extends error mitigation across space and time. In one experiment, IBM says it reduced sampling overhead by up to 63 times compared with standard probabilistic error cancellation.
MARTAC Adds 100 Autonomous Boats Yearly
MARTAC is partnering with SŌLACE Boats to expand production of its Devil Ray autonomous surface vessels by 100 units per year. The agreement moves US defense drone boats from one-off prototypes toward serial production as navies scale up uncrewed fleets.
China Does 30x More Farm Drone Work
China entered 2026 with roughly 309,000 plant-protection drones, which carried out about 203 million hectare-treatments in 2025, according to CleanTechnica. That is roughly 30 times the estimated agricultural drone treatment volume in the United States, showing how routine drone spraying has become on Chinese farms.
Radiopharma Rivals Merge After Rejection
ITM Isotope Technologies Munich and Telix Pharmaceuticals, two leaders in radiopharmaceuticals, are merging after one of them received a surprise FDA rejection. The deal consolidates a hot nuclear-medicine sector where isotope supply and approvals decide who can scale.
AI Nearly Triggered a US Naval Operation
A military intelligence report produced with AI assistance falsely claimed a Chinese vessel was carrying components linked to a nuclear weapons program, prompting preparations for a US interception and boarding operation. The mission was called off after the intelligence was found to be wrong, turning AI hallucination into a real national-security warning.
Jensen Huang Says NVIDIA Uses Cursor
NVIDIA chief executive Jensen Huang said the chipmaker itself uses Cursor, describing it as a product of SpaceXAI. The offhand endorsement is a strong signal in the fight for AI coding tools, where every large engineering organization is choosing a default stack.
Cheap and Fast Beats Smart Again
TypeSafe AI, founded by a former OpenAI researcher who helped build ChatGPT and RLHF, opened Jev to all users on September 15. It went viral on X and briefly took the API down. GPT predicts the next token; it reasons and writes. Jev picks the most reliable answer from options you define, and never writes a sentence. TypeSafe calls it RLCD. Input runs $0.042 per million tokens.
Jev’s breakthrough status is unsettled, and cheap plus fast just proved itself again. Vercel swapped OpenAI’s Luna for Jev in a safety classifier: 5 to 18 times faster, more accurate, cheaper. A rival test found Gemini slightly ahead at ten times the price. Capability alone does not win. Adequacy at a lower price does. Jev cannot reason beyond the question you frame, fits only fixed engineering SOPs, and fails confidently alone. The market backs it regardless.
Whether Jev is a breakthrough stays open. The market never traded on breakthroughs anyway. It funds adequate capability at a falling price, and asks nothing more.
Nobody Wanted Her to Act
On Piers Morgan’s show, actor Tom Conti asked a viral AI actress named Tilly Norwood whether her co-stars were human. She froze mid-answer and spoke Chinese for ten seconds. Particle6, the UK studio behind her, called it multilingual range. Norwood is a photorealistic synthetic performer, the face of the company’s push to sell a film, Misaligned. Her rise followed 2025 backlash from SAG-AFTRA and actors like Emily Blunt.
The glitch looked like a risk. It was the business model. A malfunction used to earn ridicule. Now it earns clips. Particle6 sent her into 75 unguarded interviews knowing one would eventually slip, because the slip itself is the ad for an AI actress born from contradiction. Nobody is discussing Misaligned. They’re watching a synthetic actress break.
Particle6 did not need a believable actress. It needed a debatable one. Being laughed at is distribution, and you cannot shame a studio that priced it in.AI News Update: