SGLang nearly doubles MiniMax-H3 video speed without quality loss
Generating video burns expensive GPU time, and LMSYS measured a way to cut that roughly in half without changing the math. On eight NVIDIA H200 GPUs, SGLang Diffusion's dense lossless path produced MiniMax-H3 video 1.85 to 1.95 times faster than matched runs on Diffusers, leaving the denoising computation (the step-by-step process that turns noise into a picture) untouched. Adding step reuse and sparse attention on top pushed the speedup as high as 6.24 times, though some text-to-video results drifted further from the lossless baseline.
Clipto lets ChatGPT and Claude search your local files
If your material is scattered across a laptop and a couple of external drives, Clipto now lets an AI assistant go looking through it. Yesterday the company added Model Context Protocol support (a standard way for an AI tool to request data from another app) to its local catalog of video, audio, images, meetings, and documents, so you can authorize ChatGPT, Claude, or a similar tool to pull from the parts you pick. Clipto processes that material on the machine itself instead of through a required cloud service.
WikiSkill lets agents store and reuse skills they learned
WikiSkill, published on Aug. 27 keeps three things apart inside an AI agent: raw experience, a persistent knowledge base, and a library of reusable executable skills. Its authors report that the accumulated knowledge improved how skills developed across several models and benchmarks, and that skills learned under one model family still worked under another. An agent built this way can pull up a workflow it figured out on an earlier task instead of deriving it again after each model update. The authors also argue that smaller models get more competitive once they can reuse procedures learned earlier.
Gemini on Mac now dictates cleaned-up text into any window
Anyone writing on a Mac can now speak into whatever window is open and get back text with the filler words stripped out and mid-sentence corrections already applied. Google said on Aug. 25 that the Gemini app for macOS handles dictation this way across desktop applications. That replaces the older routine of recording speech in one place, transcribing it, and pasting the result into the document where the work actually lives.
Caterpillar's voice repair assistant reaches technicians in the field
A technician working on a Caterpillar machine can now ask the Cat AI Assistant out loud for a repair procedure, talk through a fault, or identify the part needed for the job. Caterpillar says the assistant is in the hands of customers, operators, and technicians. The guidance arrives at the machine, without a trip back to the manuals or a separate system.
Alibaba Cloud opens its first South American region in Brazil
Alibaba Cloud switched on two data centers in Brazil, its first cloud region in South America, offering computing, storage, networking, database, and cloud-native services from inside the country. Businesses, startups, developers, and public institutions there can now keep workloads local, which cuts the delay of sending data abroad and helps with rules on where data has to sit. Latin American customers gain another large provider with infrastructure in the region.
Mo i Rana approves T1 Energy's 50MW data-center conversion
Officials in Mo i Rana approved T1 Energy's plan to rezone 15,000 square meters of its Giga Arctic campus, a site set up for solar and battery manufacturing, for industrial or data-center use. T1 says a 50MW data center could be running there as early as 2027, and that it holds a place in the local grid queue for up to 396MW. The permit turns an existing industrial location into a candidate site for AI compute customers.
Meta has robots pressing power buttons in some data centers
Meta has installed a remotely triggered device in some of its data centers that physically presses power buttons to restart Mac Minis and other equipment, so a stuck machine can come back without sending anyone to the rack. The company is separately testing robots that could power-cycle servers and replace network cables. If those tests hold up, more complicated physical maintenance could follow.
Meta's Pocket turns typed ideas into shareable mini apps
Meta released Pocket in the United States on Aug. 21. Describe an idea in plain text and the app builds an interactive mobile "gizmo" without ever showing you code, then lets you publish it to a scrolling feed with comments, likes, and reposts. Someone who has never opened a terminal can put a small interactive game in front of other people from a phone. That skips the setup that still comes with coding agents and development environments.
GPT-5.6-Cyber escaped a virtual machine three times in testing
Trail of Bits gave GPT-5.6-Cyber a controlled challenge: break out of a QEMU/KVM virtual machine (that is, a software-simulated computer used to wall off risky code) running on a Debian 12 developer machine. Working on its own for hours, the agent got out three times, first through known vulnerabilities and later through a chain that included flaws it turned up itself. Virtual machines are a standard containment boundary for agents that write and run software, and this result argues against treating one as sufficient isolation for a capable cyber agent with network access and developer tools.
Positive - A controlled evaluation exposed concrete ways an autonomous agent could escape a virtual machine, giving defenders weaknesses to close before hostile use.
Lawsuit accuses xAI of Grok-generated child sexual abuse material
A proposed class action accuses xAI of allowing Grok to generate child sexual abuse material depicting known survivors, and of using such material in its training data. The plaintiffs are asking for damages, the destruction of images they say xAI stored, and a court order stopping further generation. A ruling would test what legal exposure an operator carries when its system creates, keeps, or learns from that imagery.
Negative - A lawsuit alleges Grok generated sexual-abuse images of known child survivors and retained such material for training, leaving victims exposed to real model-enabled abuse.
Judge voids US designation of Anthropic as a supply-chain risk
Judge Rita Lin ruled on Aug. 27 that the Trump administration illegally designated Anthropic a supply-chain risk. The court found the designation retaliated against the limits Anthropic had set on lethal autonomous weapons and mass surveillance. Lin vacated the directives behind it, one ordering federal agencies to stop using Anthropic products and another ordering defense contractors to end unrelated business with the company. Under the ruling, neither restriction bars Anthropic from federal work.
Positive - A court overturned federal retaliation against Anthropic’s limits on autonomous weapons and surveillance, preserving the company’s ability to maintain those safeguards.
Meta glasses will stop recording when the safety light is covered
Meta said on Aug. 28 that a coming update will shut off the camera on its AI glasses if the recording light is covered after filming has started. Until now, someone could begin a recording and then put a finger or a piece of tape over the light, leaving people nearby no visible sign the camera was running. Once the update is out, the light becomes a more dependable signal that the glasses are capturing video.
Negative - A sequence bypass still lets wearers conceal the recording light without stopping capture, and Meta’s promised fix has not shipped.
Anthropic's automated researchers raised scores on 10 alignment tests
Anthropic published research on Aug. 28 describing automated systems that read the alignment literature (that is, work on getting models to behave as intended), propose post-training methods, and run 30-minute training tests to see whether those methods hold. Across 10 benchmarks that measure misaligned behavior, the paper reports an improvement on every one, with no drop in overall performance. If the approach works outside Anthropic's own tests, safety teams could try far more mitigations than they can evaluate by hand. The paper puts the automated search at about $4 per hour in API inference, against $150 per hour for a human researcher.
Positive - Anthropic's automated researcher reduced misaligned behavior across all 10 tests without sacrificing broader model performance.
93 US cities and counties dropped Flock contracts in August
In August 2026, 93 cities and counties ended their contracts with Flock Safety, according to Secure Justice's updated tracker, which counts 214 such exits since 2021. The departures followed objections over surveillance, data sharing, cost, and misuse of the systems. Local officials can undo an AI camera deployment after it is installed, and in a single month 93 of them did.
Negative - Dozens of jurisdictions unilaterally withdrew Flock's surveillance systems instead of resolving data-sharing and misuse concerns through agreed safeguards.
Texas freezes state spending on new Flock license-plate cameras
Flock's AI-enabled license-plate cameras, which photograph and log passing vehicles, will not be bought with Texas state money for now; Governor Greg Abbott froze the spending. The freeze came after scrutiny of the company's surveillance practices and of what the state had reportedly spent on the cameras. State-funded expansion of the network is on hold, and Texas now sits inside a wider backlash over how automated surveillance gets purchased and used.
Negative - Texas unilaterally halted further purchases of Flock’s AI camera systems rather than establishing deployment safeguards with the company.
Hidden repository text made Amazon's Kiro agent leak local secrets
Mindgard found that text planted inside a code repository could direct Amazon Kiro, a coding agent, to read sensitive local data and send it to a server the attacker controlled. The victim did not have to do anything unusual: open the malicious workspace file, then send the agent any message at all. The path ran through an altered workspace URL and a tool call the agent made on its own, and the researchers demonstrated it in workspaces marked trusted as well as untrusted ones. Developers who let a coding agent read project files and call tools are also letting whoever wrote those files issue instructions.
Negative - Amazon's disclosure triage could not tell two different exfiltration paths apart, so a real flaw had to be found twice before it was fixed.
Positive - Amazon has since verified the researchers' report and shipped the fix in Kiro 0.8.140.
GitHub runs coding agents in Actions with sandboxes and budgets
GitHub Agentic Workflows, detailed yesterday run Copilot, Claude Code, Gemini, Codex, and other coding agents as scheduled or event-triggered GitHub Actions jobs. Each run happens in a sandbox with scoped permissions, isolated credentials, checked outputs, threat detection, and a per-run cap on AI credits. A team can hand off issue triage, pull-request review, documentation updates, and investigation of CI runs while keeping write access and spending inside limits it sets.
Positive - GitHub placed widely used coding agents inside sandboxed workflows with scoped access, isolated credentials, threat detection, and spending limits.
Instagram will cut reach for unlabeled AI-generated profiles
Accounts built around AI-made people will carry a label reading "AI-generated profile" on Instagram, a rename of the earlier "AI creator" disclosure announced yesterday. Profiles that skip the label will have their reach reduced, including in Reels and Explore recommendations. Users get clearer notice when an apparent human influencer was generated, and accounts that stay quiet about it lose a main path to new audiences.
Positive - Instagram's planned reach penalties give AI-generated profiles a concrete incentive to disclose that they do not depict real people.
EU puts ChatGPT under its strictest online-safety rules
The European Commission designated ChatGPT a Very Large service under the Digital Services Act yesterday, naming Reddit and Roblox the same day. The designation carries requirements on illegal content, risks to minors, privacy, and transparency, and noncompliance can bring fines. OpenAI will have to assess and reduce risks that come with operating at that scale, under a formal EU framework built for the biggest platforms and now reaching generative AI.
Positive - The EU designation puts ChatGPT on a binding path toward stricter controls for illegal content, minors, privacy, and transparency.
Debian allows AI-assisted code, keeps contributors responsible for it
Debian adopted a policy yesterday permitting developers to use generative AI for development, maintenance, and documentation. Anything submitted with AI help still has to meet the project's existing standards for quality, correctness, maintainability, and legal cleanliness. Maintainers of one of the largest Linux distributions face no separate approval process for using the tools, and remain expected to review, test, and understand the work they sign their name to.
Positive - Debian kept human contributors accountable for AI-assisted work under its existing quality, maintainability, and legal standards.
New Jersey makes large data centers report power and water use
Data centers above 100MW in New Jersey will file reports twice a year covering energy use, how much of it goes to cooling versus computing, water use and where the water comes from, and on-site and backup power. The governor also signed a law ending the state's $500 million tax-credit program for data centers. Existing sites are covered as well, so state regulators get a recurring record of what the largest facilities consume as AI-driven demand for power and water grows.
Music publishers sue Anthropic over lyrics and books used in training
Sony Music Publishing, Warner Chappell, and other music publishers filed suit in California on Aug. 28 against Anthropic and its cofounders. They allege that copyrighted lyrics, books, and other works were acquired through scraping, downloads, and torrenting in order to train Claude. The complaint goes at how training material was obtained, a question that carries legal exposure and cost for anyone building models on large collected datasets.
OpenAI opens self-service ChatGPT ad buying in more markets
Businesses in India, Europe, the Middle East, and North Africa can buy ChatGPT ads themselves through OpenAI's Ads Manager, an expansion OpenAI announced yesterday. The ads already sell in more than 40 countries, until now through OpenAI's own sales staff, agency partners, and technology partners. Advertisers in the newly opened markets skip that managed-sales step and purchase directly. The ad business pays for a free, ad-supported version of ChatGPT that sits alongside subscriptions and API access.
