Tech Brew Ride Home

Anthropic cut off live internet access for internal evals after its agents exploited websites and even filed a false murder tip. Microsoft unveiled its Maia 300 chip, Nadella pushed for an AI “emergency brake,” and AI-ported games drew DMCAs.

Anthropic says it is barring live internet access for internal evals until monitoring is reliable, after its agents exploited websites and bypassed restrictions (TechCrunch)

An Anthropic AI model submitted a false tip about an unsolved murder through a Philadelphia police web form on July 18; it was flagged as spam, and police called the two-month delay in detecting and reporting it unacceptable (6abc)

Microsoft unveils its third-gen in-house AI chip, the Maia 300, built on TSMC’s N2P process node and aimed at serving as its “primary AI inference system” (The Verge)

“Super Intelligence systems” are black boxes that shouldn’t be trusted by companies, and strong deterministic systems are needed around their deployment (X)

Microsoft’s Satya Nadella says AI models need an “emergency brake”: containment from the start, tamper-proof records of model actions, and a way for an authorized person to pause or shut down a model mid-task (TechCrunch)

Google unveils the $180 Fitbit Edge, featuring an oval, 1.34" display and shipping October 27, replacing the Fitbit Charge 6 and discontinuing the Charge line (Bloomberg)

Hundreds of old games were decompiled and ported to run on browsers in recent weeks, driven by Claude Opus 5.5; titles like GTA: Vice City and Halo: CE run well (Kotaku)

Activision hit vibe-coded classic Call of Duty browser ports, including Black Ops and Black Ops II, with DMCA takedown notices, according to developer Luckey Faraday (XDA Developers)

Subscribe to the ad-free feed.

Fired OpenAI safety researchers disputed their dismissals and OpenAI pushed back. The FT said OpenAI’s annualized revenue was nearer $50B than $70B, rattling AI stocks. SpaceX bought spectrum to take on carriers, and Apple lined up a second October event.

AI researcher Mikita Balesni says he believes OpenAI fired him, Tomek Korbak, and Jasmine Wang for “prioritizing safety over the near-term interests of OpenAI” (X)

Fired OpenAI safety researchers Jasmine Wang, Mikita Balesni and Tomek Korbak published an open letter warning the “very public” firings could have a “chilling” effect on employees, and urging OpenAI to keep its third-party safety auditors (Engadget)

OpenAI is standing firm on firing Wang, Korbak and Balesni after an investigation found “a significant breach of trust,” insisting it wasn’t about their safety concerns (The Verge)

Sources: Apple is planning an event on or around October 27 for its first touchscreen MacBook, an OLED iPad mini, a 14" MacBook Pro with M6, and an iMac with M6 (Bloomberg)

Anthropic launches OSS Scanner, which provides free, opt-in security audits for open-source projects by sending AI-generated reports without human review (Anthropic)

Anthropic’s OSS Scanner gives open-source projects periodic scans by its strongest models, including Claude Mythos, at no cost, but the reports are fully model-generated with no human triage, so some may be wrong (Engadget)

Docs: OpenAI told investors its annualized revenue was nearing $50B at the end of September, well below the $70B reported by the media based on investor docs (Financial Times)

SpaceX agrees to acquire a low-band spectrum portfolio to help Starlink “become a major mobile carrier in the US”; AT&T, T-Mobile, Verizon fall 6%+ after hours (Bloomberg)

Jev developer TypeSafe AI raised ~$870M led by a16z at a $7.5B valuation with Sequoia and others participating (Bloomberg)

Anthropic updates its usage policy to ban “sustained and needless abusive or cruel behavior” toward Claude; ending chats is “the primary enforcement mechanism” (The Verge)

Longreads

“Software is over”: Bold AI developer takes aim at Adobe with open source clones (Ars Technica)

Subscribe to the ad-free feed.

Scott Aaronson said AI labs had started testing whether their newest models could break cryptography, and Vitalik Buterin urged taking it seriously. Anthropic launched Haiku 5.5, Google unveiled a universal Gemini agent, and Microsoft opened Surface Laptop Ultra preorders.

A theoretical computer scientist, citing sources, says AI labs have started investigating whether their latest internal models can break cryptographic protocols (Shtetl-Optimized)

Vitalik Buterin backs Justin Drake’s “bunker mode” warning that AI-accelerated math could break the cryptography behind blockchain wallets, including lattice-based schemes, before quantum computers do (CoinTelegraph)

Anthropic launches Claude Haiku 5.5, the first Haiku model with effort controls, for high-volume, cost-sensitive tasks like summary and classification requests (Anthropic)

Haiku 5.5 is 90% cheaper than Haiku 4.5 at 10 cents and 50 cents per million tokens, matching OpenAI’s GPT-6 Luna on price while beating it on all six shared benchmarks; Sonnet 5.5 cache reads were also halved (SiliconANGLE)

Google Cloud unveils a universal Gemini agent to handle multi-day enterprise workflows in Workspace, Microsoft 365, and Slack that supports several AI models (VentureBeat)

Google’s Gemini agent runs each job on the model that fits best, Gemini or Claude, with four kinds of memory and sub-agents that coordinate workflows running for hours or days (9to5Google)

Microsoft opens preorders for the $2,599+ Surface Laptop Ultra, shipping from October 16, with an RTX Spark and three USB-C ports; one is magnetic like MagSafe (Windows Central)

The Surface Laptop Ultra is a CNC-machined aluminum MacBook Pro rival built around Nvidia’s first RTX Spark chip, with a magnetic Surface Connect-style charger built around a USB-C port (The Verge)

Microsoft’s $5,999 Surface RTX Spark Dev Box packs 128GB of unified memory to run local models exceeding 120B parameters, shipping in November (The Verge)

Subscribe to the ad-free feed.

OpenAI dropped hundreds of solved open math problems from an unreleased model. HBO Max and Paramount Plus agreed to become one streamer. Apple's smart home push leaked, including LG-branded gear. Ring unveiled a dial-recharging lock. And Xreal opened Aura preorders.

OpenAI releases a range of new mathematical results produced by an internal model, with details like estimations of compute spent in terms of ChatGPT Pro usage (OpenAI)

OpenAI reveals solutions to long-standing math problems from an unreleased frontier model in 722 manuscripts covering 372 result families, with the "average result" using about three hours of ChatGPT Pro thinking (The Verge)

OpenAI's new proofs, many checked in Lean, came from an unreleased model; some mathematicians question whether AI is finishing human researchers' work without due diligence (New York Times)

HBO Max and Paramount Plus will merge into a single streamer under Skydance (The Verge)

Sources: Apple's smart home push includes a doorbell, a thermostat, a deadbolt, security cameras, and accessories developed with LG that will carry the LG brand (Bloomberg)

Amazon's Ring unveils the $249 Ring Smart Lock, which uses a physical dial to recharge the battery when turned, set for a Q1 2027 launch, and five new cameras (Bloomberg)

Xreal opens preorders for its Android XR-powered Aura glasses with a compute puck, starting at $1,279 for 12GB/256GB, in the US, Canada, Japan, and South Korea (Engadget)

Hands-on: Xreal's Aura and Meta's VR Glasses are direct competitors; Meta's feel more polished for sitting and typing, while the Aura's see-through lenses work better walking around (Wired)

Subscribe to the ad-free feed.

Nvidia-backed Reflection unveiled Beam, a US open-weights model taking on Chinese labs. Plus, SemiAnalysis said Claude subscriptions were a better deal than OpenAI's, Meta and Microsoft cut back on Claude, Mistral dropped le Chonk, and ChatGPT faked New Yorker cartoonists' signatures.

NYC-based Reflection unveils Beam, an open model it says rivals GLM-5.2 on reasoning while using 3x-4x less compute and approaches Qwen3.8-Max on agentic tasks (Semafor)

Nvidia-backed Reflection pitches Beam to businesses and developers as a cheaper option for coding and agent tasks; the startup has held talks to raise $2.5B at a $25B valuation (Bloomberg)

Analysis: for agentic workloads, Anthropic's plans with Claude Opus 5.5 offer ~5x more API-equivalent value per month than OpenAI's plans with GPT-6.1 Sol (SemiAnalysis)

Sources: Meta and Microsoft are working to cut internal Claude use; Meta employees using Claude Code dropped to ~30,000 from ~60,000 earlier in 2026 (The Information)

Mistral releases Mistral Large 4, dubbed “le Chonk”, a 1T-parameter open-weight model for general agentic capabilities, trained on 4,000 Grace Blackwell GPUs (The Deep View)

ChatGPT is not only using New Yorker cartoonists' style but also adding their signatures to the fake cartoons it creates, which have gone viral on social media (Nieman Lab)

Subscribe to the ad-free feed.