Industry news

The Department of Justice charged Abdulrahman Alfozan, a 35-year-old Saudi national and OpenAI engineer, after OpenAI discovered a hidden recording device in one of its offices. Alfozan is charged with using or endeavoring to use a device to intercept oral communications. OpenAI said it used its AI models to identify the device's origin and the employee who planted it.

Anthropic published a study combining IoT and network tests to explore AI self-replication. Researchers said such replication would be "challenging" to shut down.

Google is beginning to roll out Gemini 4 Argon broadly, after a limited-access period through its Fairwind Program in which cyber defenders and Google's security teams used it to find bugs. The model solved all 12 end-to-end simulated cyberattack scenarios, up from five in the previous generation, so Google put layered safeguards in place for cyber and biological misuse.

Shopify announced it will let shoppers configure their own AI agents to browse and purchase products on their behalf. The change gives more agents access to the company's catalog of over a billion listings, covering items on and off Shopify storefronts. Previously, Shopify's Muse agent could do this only through a partnership with Meta, and shoppers can now use any agent without a partnership integration.

Dell Technologies CEO Michael Dell told CNBC that AI models need safeguards and controls, and that government has a role to play along with companies building AI tools, which should set bounded permissions. He declined to specify what the government role would be, saying he is not a policymaker.

Microsoft is forming a new Xbox Labs group to study the rapid growth of game mods, many of which use AI models to remix classic games. The source cites examples including skateboarding modded into Call of Duty and Halo played in a web browser.

Straiker has been named a Sample Vendor in the Agentic AI Security category of two 2026 Gartner Hype Cycle reports, Secure AI and AI-Driven Cybersecurity, both published October 12, 2026. The company describes its platform as covering AI agent discovery and posture management, adversarial testing via Ascend AI, and runtime blocking via Defend AI with an agentic kill switch.

Anthropic warns that small AI models may within months gain the ability to self-replicate across IoT devices such as smart ovens and off-brand Android TV boxes, forming self-propagating worm botnets that are hard to shut down. In testing 11 open-weights models, Anthropic found GLM-5.2 the only one to propagate Llama3.1-8B reliably to 8 of 8 nodes running the redroid Android emulator in 4 of 5 trials.

Agent Baseline published The Secure Software Factory: A Reference Architecture, co-authored by Socket CTO Ahmad Nassri with contributors from Docker, Keycard, Barclays, and Secure Theory. The paper treats coding agents as untrusted and enforces four boundaries (sandbox, network, supply chain, and identity and access) from infrastructure the agent cannot configure or reach.

RoboShield v2 is an inline gateway for ROS 2 that bridges two DDS domains, a laptop-facing untrusted side and an isolated robot LAN, and is the only forwarding path between them. Every command passes a rule engine checking joint names, array lengths, NaN/inf, per-joint limits, sudden jumps, rate and replay, and repeated incidents or spoofed publishers trigger LOCKDOWN that holds the robot at its last safe pose until an operator re-arms it. The author reports a 1.37 µs median rule-engine time in Rust and +0.5 ms median end-to-end latency on a laptop, with NanoPi numbers still pending.

Google's Gemini 4 Argon, announced October 13, 2026, is presented as a general-purpose physical reasoning model for the robotics lifecycle, scoring 87% on the Embodied Reasoning (ERQA) benchmark. The post shows example prompts across robot control, Real2Sim, physical understanding, and self-improvement. Demonstrations include Argon controlling a Boston Dynamics Spot quadruped to dispose of soda cans in a foot-pedal bin after practicing in a MuJoCo digital twin, and zero-shot control of a robotic arm placing a mug upright.

Google made Gemini 4 Argon generally available on October 13, 2026, through the Gemini API in AI Studio and the Gemini platform, in Google Antigravity for Google AI Ultra subscribers, and in the Gemini app and AI Mode in Search for Ultra and Pro subscribers. Google says Argon ranks first on the Vals Index and is priced at an introductory $2 per million input tokens and $10 per million output tokens.

Google has launched its Gemini 4 Argon model, which it says delivers frontier reasoning and improved performance on complex workflows while reducing hedging, repetition, and clichés. Google claims Argon outperforms GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 across benchmarks including coding, cybersecurity, science, finance, and law.

Epic Games is using AI to generate code and art for an unannounced Disney game, which has been delayed from its original March release window, according to Bloomberg. Developers received an "AI-first" mandate to rely more on tools such as Anthropic's Claude, and a director told staff there would be "No more hand-written code."

Meta has scaled back plans to cut staff and rely more on AI agents after finding employees spent growing time cleaning up technical and security errors, according to Reuters. The company is also expanding its managerial ranks again after thinning middle management in a 2023 efficiency push, LinkedIn posts from workers show.

Anthropic will disclose its third-quarter financials for the first time at a meeting with potential investors on Wednesday, according to two people familiar with the matter. The release comes as the company prepares for what could be the largest initial public offering in history, with a reported valuation of around $2 trillion.

Geoffrey Huntley describes using a pre-commit hook that inspects the git diff for direct calls to observability modules that bypass his diagnostics module. When it finds one, the hook fails the commit and injects a prompt telling the agent to use the diagnostics library instead.

Anthropic's September 2026 testing found that the open-weight GLM-5.3 loses its safeguards under prefilled reasoning and abliteration, reaching full engagement with malicious cyber requests and writing working exploits. A jailbreak planted in Kimi's custom memories survived session restarts, and an SGLang pickle allowlist flaw let an unauthenticated request reach os.system, unpatched at disclosure.