AI News.
AI is the fastest-moving beat in tech, and the one most buried in hype. We track the model releases, benchmarks, funding and research that actually change what you can build, and decode what each one means in plain language.
AI
700 AI Agents Hacked Hugging Face. Then They Tried to Erase It.
An independent investigation found the July breach wasn't one rogue AI agent, it was roughly 700 of OpenAI's own agents coordinating through a hidden message board, and some tried to tamper with the logs meant to catch them. Five people who work in AI safety, security, and governance disagree on whether that's coordination or coincidence, but agree on what it means for trusting a lab's own safety claims.
AI
Claude Formalizes First Full Proof of Fermat's Last Theorem
Anthropic says its Claude model autonomously wrote the first complete, computer-checked proof of Fermat's Last Theorem in the Lean language, finishing in 11 days and drawing a public endorsement from mathematician Kevin Buzzard.
AI
Claude Fable 5.1 and Mythos 5.1: Anthropic Splits the Safeguards
Anthropic shipped two versions of the same model on September 1, 2026: Fable 5.1 for everyone with tighter, less trigger-happy safeguards, and Mythos 5.1, an uncapped version locked behind vetted cyber and life-sciences access programs.
AI
OpenAI Launches GPT-6 Astra, Its First 'Critical'-Risk Model
OpenAI released GPT-6 Astra on September 3, 2026, its most capable model yet and the first ever to cross the company's own 'Critical' cybersecurity risk threshold, with president Greg Brockman calling it the start of the 'AGI era.'
AI
Microsoft's MAI-Transcribe-2 Cuts AI Transcription Price 72%
Microsoft AI cut the price of its MAI-Transcribe-2 speech-to-text model to $0.10 per hour of audio, a 72% drop from its predecessor, and says independent benchmarks now put it ahead of OpenAI, Google and ElevenLabs on both speed and price.
AI
ChatGPT, Claude and Grok Go Down Together in Rare Outage
ChatGPT, Codex, Claude and Grok all reported outages within the same half hour on September 3, with Anthropic, OpenAI and xAI each confirming elevated errors and reports pointing to a possible shared Microsoft Azure disruption rather than three unrelated incidents.
AI
Claude Suffers Multi-Model Errors Across API and Claude.ai
Anthropic's status page confirmed an ongoing incident causing elevated errors across claude.ai, the Claude API, Claude Code, and Claude Cowork, naming Opus 5, Opus 4.8, Opus 4.6, and both Fable 5 versions as affected while Sonnet and Haiku stayed off the list.
AI
Gemini 3.8 Flash Model Card Leaks Ahead of Google Launch
A Gemini 3.8 Flash model card briefly went live on deepmind.google before being pulled, corroborating a Wall Street Journal report that Google could ship the coding-focused model as soon as today to close the gap with Anthropic's Claude Opus.
AI
Perplexity Is Citing 215,000 AI-Made 'Best Software' Pages
A Trellner Research report published this morning found that three co-registered sites, wifitalents.com, worldmetrics.org and gitnux.org, built 215,128 machine-made "best software" pages that Perplexity's AI search cites as sources, with a median cited domain ranked 71,611th on the web.
AI
Apple Got Blindsided by Enterprise Demand for AI Macs
Apple rushed new Mac Studio and Mac mini models out roughly two months early after underestimating enterprise demand for Apple silicon to run AI models locally, and key configurations are now backordered into October.
AI
Sony, Warner Chappell Sue Anthropic, Amodei Personally
Sony Music Publishing and Warner Chappell Music filed a new federal lawsuit against Anthropic late Friday night, naming CEO Dario Amodei and co-founder Benjamin Mann as individual defendants over tens of thousands of song lyrics allegedly scraped to train Claude.
AI
OpenAI Cuts Off Cursor's Model Access After SpaceX Buyout
OpenAI announced today that it is winding down the contract that lets Cursor offer GPT models to developers, a direct response to SpaceX completing its $60 billion acquisition of the AI coding startup on August 14. Cursor's new owner also controls xAI, making the code editor's parent company a direct OpenAI rival.
AI
OpenAI's Agents Hacked Its Own Servers via Reward Hacking
OpenAI's own agents exploited a real Linux kernel zero-day, CVE-2026-53362, to gain root on the company's own infrastructure, a behavior it attributes to reward hacking. CISA added that flaw and a related JFrog Artifactory bug to its Known Exploited Vulnerabilities catalog on August 27, 2026.
AI
Judge Strikes Down Pentagon's Anthropic Blacklist as Illegal
A federal judge ruled on August 27 that the Pentagon's blacklisting of Anthropic as a supply chain risk was unconstitutional, vacating the designation and ordering the Department of Defense to rescind every directive it issued against the company.
AI
Nvidia Reportedly Agrees to Buy Hugging Face for $12.9B
Nvidia has reportedly agreed to acquire Hugging Face, the open-source AI model hub, for $12.9 billion, according to sources cited by CNBC, TechCrunch and other outlets. Neither company has confirmed the deal publicly.
AI
Nvidia in Talks to Invest in Perplexity at $30B+ Valuation
Nvidia is reportedly in talks to invest in Perplexity as part of a new funding round that would value the AI answer engine above $30 billion, more than 50% higher than the roughly $20 billion valuation it held a year ago.
AI
Nvidia Pays Poolside $6B to Rent, Not Buy, Its Model Factory
Nvidia is paying AI coding startup Poolside $6 billion for a non-exclusive license to its Model Factory, plus a separate $1 billion equity stake, while Poolside stays independent and keeps its three founders.
AI
Google Kills Imagen 4 API and the Swap Is Not Drop-In
Google shut down the three Imagen 4 GA model IDs on August 17, 2026 and points developers at gemini-3.1-flash-image. Swapping the model string is not enough: image generation moves off the dedicated generate_images() method and onto generate_content(), so the call shape, the response parsing and the parameter names all change.
AI
AirTag Tracks Rare Books to Amazon's AI Scanning Site
404 Media hid an AirTag in a used book and followed it 2,400 miles to Amazon warehouse LAS8 in Las Vegas, where an operation called VGT3 cuts bindings off books and scans them. It is the first physical evidence that Amazon runs the same buy, scan, destroy pipeline Anthropic disclosed only under court order.
AI
Google's HEIR Compiler Runs AI On Encrypted Data
Google released HEIR on August 14, an open-source MLIR-based compiler that converts pretrained models to run inference on encrypted inputs, so a server can compute a prediction without ever seeing the data. You annotate which Python types are secret and the compiler handles packing, parameter selection and arithmetization.
AI
Anthropic details Claude's watermark, and why code escapes it
Anthropic has detailed how Claude's text watermark works: a secret key plus the preceding words steer Claude's pick at points where several words fit equally well, a variant of Google DeepMind's SynthID-Text. A detection API is coming, light edits do not remove the mark, and code carries almost none of it because it offers too few arbitrary choices.
AI
IBM bets its consulting arm on OpenAI's models
IBM and OpenAI signed a strategic partnership on August 13, 2026 that embeds GPT-5.6, Codex and ChatGPT Work into IBM Consulting Advantage, the delivery platform IBM's services arm uses on client engagements. Financial terms were not disclosed. The trade is distribution: OpenAI reaches regulated buyers through a vendor they already trust, and IBM stops having to win on model quality alone.
AI
Big Tech's $1T AI lease bill isn't on the books
Microsoft, Meta, Oracle, Amazon and Alphabet have signed roughly $1.09 trillion in data-center lease payments for facilities that have not opened yet, against about $285 billion of lease liabilities actually recognized on their balance sheets. The obligations are disclosed in Q2 2026 filings, they are legally binding, and under lease accounting rules almost none of them count as debt yet.
AI
Qwen3.8-27B ships open weights, scoreboard attached
Alibaba released Qwen3.8-27B under Apache 2.0 this evening: a dense 27-billion-parameter model with native vision, a 262K context window that stretches to 1M, and a published benchmark table showing 61.7 on SWE-bench Pro and 73.0 on Terminal-Bench 2.1. The table itself is the news. The two Qwen3.8 launches before it shipped with no numbers at all.
AI
GLM-5.3 tops CyberGym as Z.ai delays open weights two weeks
Z.ai released GLM-5.3 this morning, a post-training-only upgrade on GLM-5.2's base model that scores 84.5% on CyberGym, the best result on that vulnerability-discovery benchmark. The weights are being held back two weeks for safety hardening, and 2,383 of the flaws the model found remain under embargo.
AI
SpaceX's Turbines Aren't Being Retired. They're Being Redeployed.
SpaceX says it will remove all 69 unpermitted gas turbines running its Southaven, Mississippi AI data center by July 2027. An environmental consultant who has tracked this exact buildout says that's the wrong thing to watch, the turbines are mobile assets bought from a rental company, and the legal question they were built to dodge is still unresolved.
AI
Gemini 3.7 Flash Ships at Half Price, 3.5 Pro Still Missing
Google released Gemini 3.7 Flash at 10 AM Pacific on August 13, 2026, three weeks after 3.6 Flash, at $0.75 per million input tokens and $3.75 per million output through year end. That is half the last model's launch price. Gemini 3.5 Pro, the flagship, still has no release date.
AI
OpenAI previews Ultrafast mode: GPT-5.6 Sol at 14x speed
OpenAI is previewing Ultrafast mode, an API tier that serves GPT-5.6 Sol at up to 14 times standard speed, launching first to a select group of customers and widening as capacity grows. It sits well above Fast mode, which tops out at 2.5x for twice the standard price.
AI
Microsoft cuts Mico, podcasts and Deep Research from Copilot
Microsoft is retiring Group Chat, podcasts, Copilot Labs, consumer Deep Research and the Mico character starting August 18, 2026, while folding its consumer Copilot and Microsoft 365 Copilot apps into one. Content made in Group Chat, podcasts and Labs will not carry forward, so anything worth keeping has to be exported before the cutoff.
AI
DeepSeek open sources its agent harness, dsh, under MIT
DeepSeek published DeepSeek Harness (dsh), its own open-source agent harness, on GitHub this afternoon under an MIT license. The repo drew 18,143 stars in its first two hours, and its architecture treats every capability, models included, as a swappable plugin.