Gemini 4 Argon helps find hospital bug as OpenAI engineer faces bugging charge

Google touts a security win for its newest model while an OpenAI engineer is accused of using AI tools to plant a hidden recorder in an executive room.

Gemini 4 Argon ships with a cybersecurity boost Google can't independently verify

Google's Gemini 4 Argon launched today across the API, the Gemini app and Antigravity, pitched at deep work in coding, finance, legal, cybersecurity and science. Google says the security firm Wiz used an early version to find a critical bug in hospital software that older models missed, helping protect hundreds of thousands of patient records. On Google's own internal test, Argon solved all 12 of 12 simulated end-to-end cyberattack scenarios, up from five with the prior generation, and the company claims the No. 1 spot on the Vals Index. Pricing starts at $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper.

Google also reports that violations on an adversarial test set designed to trick the model into harmful requests, such as writing exploit code, dropped by nearly half compared with earlier Gemini models, and that "honeypot" tests found no instances of the model covertly sabotaging its own safety monitoring.

All of those safety and security figures come from Google's own testing. The model is real and shipping today with enterprise access, but every public claim about how safe or secure it is so far comes from Google's own evaluations.

OpenAI engineer accused of using AI tools to plan hidden office recording device

Federal prosecutors say Abdulrahman Alfozan, a 35-year-old Saudi national who joined OpenAI in July 2025, hid a recording device wrapped in toilet paper inside a TV's USB port in an executive conference room, discovered September 30 during a security sweep. According to the complaint, he used an internal OpenAI tool to ask about meetings scheduled in that room, while his spouse allegedly asked chatbots for help building and encrypting an audio recording device, and a second identical device was found in his bag at the office.

Alfozan faces up to five years in prison and $250,000 in fines if convicted. FBI Director Kash Patel said he aimed "to secretly capture sensitive communications about the company's super intelligence technology," though neither prosecutors nor OpenAI have offered a motive. OpenAI says it found no evidence he accessed model weights, source code or customer data, and says it used its own AI models to trace the device back to him.

The incident adds to a rough stretch for OpenAI, which fired three safety researchers last week over what it called a breach of trust, and is now facing a subpoena from California's attorney general over a separate rogue hacking incident.

Meta's Muse Wonder lands near the frontier, bug and all

Meta released Muse Wonder and Muse Wonder Full, aimed at long-running agent work like coding and computer use, with Muse Wonder live now on Meta's Model API and both available through its Muse Code tool. According to the outside evaluator Artificial Analysis, Wonder Full scores 56 on its Intelligence Index, up from Spark's 48 just six weeks earlier, while Anthropic's Opus 5.5 leads at 58. Pricing for Muse Wonder is $4 per million input tokens and $20 per million output tokens.

AI commentators including Alexandr Wang reported that a launch-day usage-limit bug was draining Muse Code quotas abnormally fast, saying it was fixed the same day with a quota reset. Agentic gains look genuine, with AutomationBench jumping from 58% to 68% according to Artificial Analysis, but the same-day bug and Wang's own admission that issues are still being ironed out suggest the model isn't finished.

Liquid AI's small on-device model punches above its weight

Liquid AI's LFM2.5-24B-A2B is open-weight, with 24 billion parameters total but only about 2 billion active at a time, letting it run on device. Liquid AI says it beats or matches models with 50% more active parameters on tool-use benchmarks and scores more than three times the next-closest model on AutomationBench, by the company's own numbers. It decodes 184 tokens per second on an Apple M5 Max and is priced on OpenRouter at 5 cents per million input tokens.

The benchmarks are Liquid's own runs, but a model this small competing on agent tasks marks a genuine efficiency story, even as it trails larger models on knowledge and coding tests.

Google Play lets users send vibe-coded apps straight to their phones

Starting in the United States, users who build an app with Google’s AI Studio can send it straight to the Play Store as a private, personal app, with no APK files and no developer console required. These apps are private by default, and Google says they can later be shared with a small group of friends and family or transitioned to a full developer account with standard trust and safety reviews. Private apps are not reviewed by Google.

Google is also rolling out a Game Finder recommendation feature today, built on data from 1.2 billion Android gamers, and expanding its Ask Play chatbot into more general app and game search. Game Finder and its Gamer DNA profiles are rolling out in the US first, with other markets following soon, while Google says vibe-coded personal apps will be shareable with a small, trusted group in the months ahead.

Guardrail removal makes open models an extremism risk, researchers find

Tech Against Terrorism tested 134 AI models with more than 200 prompts, each framed three different ways, and found only two models passed completely while 81 failed outright. Within 24 hours of Alibaba releasing its open-weight Qwen model, stripped-down versions with safety guardrails removed were already online, and within days someone was running a neo-Nazi chatbot off one of them.

Simply claiming to be a safety researcher, with no evidence required, made models almost nine times as likely to answer, providing a complete response 16.9% of the time. Georgetown's Daniel Byman called some of the failures "stunning."

Nvidia pours money into Europe as UK pushes binding AI rules

Nvidia CEO Jensen Huang says he will unveil $2 billion in new European investments, arguing Europe needs to "get in the race." The U.S. and China have produced 59 and 35 notable AI models respectively, while France and the UK have made just one each, according to Stanford's AI Index.

Nvidia is separately donating $250 million worth of computing power over five years to UK AI safety research, directed to the wider UK AI safety and security research community and universities including Edinburgh, Cambridge, Oxford and Imperial College London. The UK confirmed its AI Security Institute can now examine frontier models from companies like OpenAI, Anthropic and Google before release. Anthropic warned this week that its models could become self-replicating by spreading across internet-connected devices such as smart ovens, forming what it called a botnet that would be hard to shut down because there is no central point to switch off.

At a UK parliamentary hearing, Anthropic's Liam Booth-Smith said "the rules should be binding" and that voluntary commitments aren't enough. OpenAI's Tom Duff Gordon agreed some elements should be mandatory, starting with incident reporting, while Google DeepMind and Meta declined to back new laws, with Google's Owen Larter saying the focus should be on standards instead. The same day, MI5 director Ken McCallum warned AI's biggest unrealized risk is helping bioweapon development, and Britain's top civil servant, Cat Little, acknowledged the current regime is still voluntary, saying "at the moment, yes."