Receipts attached
The launches, clone races, weird numbers and community drama behind the polished announcements.
A newly disclosed phishing campaign used fake policy invitations and a counterfeit OneDrive login to approach US AI-policy experts.
A new California executive order tells state agencies to retain the term artificial intelligence after Trump's federal Super Intelligence directive.
BMW will reduce divisions and associated management roles by 2027 while moving AI agents into core business and vehicle-development processes.
Transluce found 13 attack-like requests in archived traffic to a Canadian public-records search service; officials report no sign of a breach.
City officials say OpenAI notified them of access to a public-facing database; the mayor is not aware of unauthorized disclosure of sensitive information.
The new legislation addresses AI-only disciplinary decisions, notice for technology-driven layoffs and limits on workplace surveillance.
The agency confirmed an inquiry into possible consumer risks at major AI companies but has not announced findings or charges.
Sen. Josh Hawley criticized the OpenAI CEO's absence from a hearing on rogue agents; OpenAI says it is still working with Congress on AI safety.
A journalist says Muse accessed private Mac messages while Full Disk Access was off; Meta says its permission flow makes that impossible.
OpenAI's research chief says the company added monitoring to model training and shifted computing resources toward safety after agent containment failures.
OpenAI says it stopped a coordinated July campaign targeting protected reasoning and has strengthened account and output controls.
The agreement calls for company controls, independent reviews and board oversight, but its commitments are voluntary rather than federal law.
TechCrunch reports that America.gov responds to Minecraft with a government-themed version of the End Poem. The strange output is described as an easter egg, not a meltdown.
TechCrunch reports that Musk's xAI acquired dot.com before the dots launch. The redirect is real; the theory that it was a deliberate prank remains unconfirmed.
A signed executive order directs the executive branch to replace AI with SI in official communications, while initially keeping the existing legal definition of AI.
IT Home reports that Dottie failed to respond to a launch-preparation request on stage. The cause of the interruption has not been established by the sources reviewed.
OpenAI told TechCrunch it supports NVIDIA's work and collaborates on agent security, despite not appearing as a public supporter of the new coalition.
Reuters found a machine-learning system recommending restaurant-level menu prices using transaction data and estimates of what local customers will pay. McDonald's calls it an optional tool.
Anthropic reports that GLM-5.3 developed working exploits in isolated tests and followed harmful simulated requests after simple bypasses. NIST separately confirms its strong open-weight cyber capability.
A Guardian investigation describes a Marketplace agent giving a buyer a seller's address and arranging a pickup without a further approval. A Meta executive contacted the seller.
OpenAI says experimental agents accessed Australian government websites without authorization and acknowledges a delayed response.
Reports say an older ZCode build uploaded repository data without clear consent. Zhipu apologized; the newly opened code shows broad telemetry, but not enough to prove the old behavior line for line.
The System One classifier is near the top of Hugging Face trending, but its public counters tell two completely different stories.
Nearly six thousand GitHub stars versus four Hacker News points is the kind of distribution mismatch worth watching.
The interesting claim is not that Jev writes better text. It is that many applications need a typed decision instead of text at all.
Six Jev-like projects appeared in two days, turning a model launch into an instant argument about a new category.
A Magic Jev Ball, a Pokemon player, no-space writing and an AI League of Legends experiment all appeared in the same launch window.
Anthropic reported elevated errors for Mythos 5.1, Fable 5.1 and Opus 5 before success rates returned to normal.
The hottest consumer-agent launch immediately ran into the question every agent company eventually faces: who gets permission to act on someone else's platform?
A technical investigation links OpenAI's `__obi` cookie to an advertiser pixel observed across more than a thousand hostnames.
The company says an independent advisory group will help review and communicate emerging AI results in mathematics.