AI Daily Digest

Three top stories follow an xAI pollution case that goes on, a call to rewrite how mathematics is judged, and hikers led astray by a chatbot.

Five top stories follow a lifted Claude cap, a smaller reported OpenAI round, a live-mouse bench injection, Apple Intelligence in China, and shopping agents that still book little.

Five top stories and twelve briefs follow a backdoored coding extension, a shared opt-out signal, a market turn against AI spending, and tighter safety answers.

Five top stories and six briefs follow open weights for a trillion-parameter model, a rewritten model spec, a personal agent on a Pentagon network, and a voice-AI acquisition.

Five top stories and eight briefs follow a staged frontier-model release, Azure and search earnings, team-level AI spending, and local resistance to data centers.

Five top stories and seven briefs follow a world model, unpublished scientific results, drug-development claims, new compute, and AI inside public decisions.

Five top stories and six briefs follow a listing filing, a cheaper computer-use model, code migration, GPU financing, and agents pitching newsrooms.

Five top stories and seven briefs follow company returns, community maintenance, search traffic, data-center siting, and robot faces. One open collaboration and one related paper keep earlier dates.

Five top stories and six briefs follow video-app consolidation, personal assistants, retrieval, advertising disclosure, and agent training. Two developer notes and one related paper keep their original dates.

Five top stories and nine briefs follow a shared skill standard, agent architecture, grid flexibility, operations, and a 10-meter forest map. Seven briefs are marked 48-hour catch-up.

Five top stories and eleven briefs follow personal assistants, security queues, shopping permissions, a local coding model, and the extra coordination managers report.

Five top stories and twelve briefs follow protein experiments, biotech workflows, speech models, shared-link exposure, and the limits of permission and deployment.

Five top stories and twelve briefs track creative workflows, personal assistants, chip software, industrial AI, infrastructure, platform responsibility, and compute financing.

Power constraints, tactile robotics, AI animated dramas, retail pricing, and Anthropic’s public-benefit proposal connect model capability to everyday business.

DeepSeek updates its model and search tools, while Anthropic examines changing work; website access, default models, and human takeover shape real agent workflows.

RTX Spark PCs, the open Beam model, and continuous bot oversight; permission boundaries, multilingual evaluation, and long-task context put capability claims to practical tests.

Independent-evaluator access, code review, time-series forecasting, and scientific agents; two conversations examine information gaps and physical delivery.

Election safeguards, persistent agents, on-device scam detection, and safety research computing; two conversations and coding-boundary research examine when to pause and who remains accountable.

Wider Gemini access, purchasing-intent credentials, an AI companion platform’s bankruptcy, and a regulatory hearing; two conversations and hidden-capability research examine the boundaries of trust.

Maia deployment plans, Muse memory, medical research and EU incident governance; two conversations connect expanding capabilities with reliable delivery, followed by routing systems and context tools.

Voice assistants’ execution barriers and local agents’ operating conditions; Gates and Fei-Fei Li conversations, defensive decoys, environment-side evidence records, and Spool.

Isolation in internal evaluations, independent controls, Huxe disclosure, and personal-agent privacy; interviews on readable reasoning and scoped permissions, stopping decisions, skill supply chains, and an alpha harness update.

Agent permissions after a false police tip, multimodal decision models, and AI disclosure in publishing; conversations on trust and goodput; evidence-based oversight, test reliability, and platform transitions.

GPT-6’s interactive interface, Haiku pricing, and local agents; watermark detection and teen protections; conversations on open safety tools and incident transparency, result verification research, and structured decision models.

Personal-agent protocols and tiered security access; Mistral Large 4 preview and contract-workflow evaluations; conversations on credential isolation and automated defense, browsing research, and local multimodal retrieval.

OpenAI text watermarks and visual advertising, Wikimedia’s agent-access investigation; conversations on website costs and production verification, PoS explicit states, and Dust training research.

Google pauses some open-source vulnerability submissions; ThinkingBox checks business outcomes; Altman’s new interview alongside earlier safety views; VISTA visual memory and verifiable open-source tests.

A safety dispute at OpenAI, Anthropic's engineer training program, and a reporter's Dots trial; plus robots that know when to stop, device SDKs, and decision models.

GPT-6 development guidance, a forthcoming DGX Spark configuration, transaction verification, and an orbital experiment; an RLM interview, tool evaluation, and AstaBrief add context.

Barclays expands Claude adoption, Claude Code opens up mods, and an economics essay explores execution; NVIDIA examines AI factory returns, while interviews add tool-design and financing perspectives.

Gemini 4 Argon, protected-reasoning extraction reports and voluntary frontier commitments frame a debate over capabilities and oversight.

Dots and GPT-6.1 Sol extend persistent execution; permissions, honest failure reporting and tool design become central questions.

Sonnet 5.5, external agent safeguards and AMD’s World Labs agreement connect model releases with deployment controls and infrastructure.

Official OpenAI incident reports distinguish detection, shutdown and third-party impact; auditable evidence matters beyond final model answers.

Weekend catch-up examines agent preferences, protected memory and verifiable tasks, retaining publication dates and reading limits.

Claude plugin submissions, Codex-built sales demos and UNESCO’s AI course show how tools move into concrete work.

Claude personal connectors in Slack, Gemini’s live avatar and faster local vision-language decoding reveal the conditions behind new capabilities.

Altman’s UN briefing, Claude enzyme research and Gemini text-to-speech connect governance, scientific applications and voice tools.

GPT-6 Sol and Luna, data-breach reporting and Cursor rollouts lead the issue, alongside autonomous-agent reliability research.

Grok 4.7 and Claude Opus 5.5 headline a historical catch-up, with legal, company and research updates retaining their original dates.

California, the UN and markets pursue distinct oversight tracks while StepFun releases a preview; source coverage remains incomplete.

An antitrust complaint challenges coordinated slowing, Google reports evaluation access to real companies, and experts demand independent oversight.

Anthropic appoints embedded assessors and California sets an oversight timetable; collection remains incomplete at fourteen of eighteen sources.

Anthropic publishes internal R&D metrics alongside the royal summit, specialized legal tools, sponsored-agent tests and chip roadmaps.

OpenAI publishes misalignment reports; the House passes an electricity-cost bill while a shutdown-switch fast track is blocked.

OpenAI supports key independent-evaluation provisions while reported financing, ByteDance costs and new model designs advance.

Trump rejects pacing appeals, a Scottish summit seeks principles, and earlier infrastructure transactions receive fuller source checks.

Standards-body talks, OpenAI’s IPO delay and contrasting company financial reports bring safety coordination into institutions and markets.

Amodei proposes embedded assessors and coordinated pacing as mathematicians, researchers and policymakers debate AI risks.

IPO reports, agent investigations and delayed release timelines test the distinction between plans and completed events.

Agent platforms expand while cybersecurity incidents and governance proposals make permissions and oversight more consequential.

Coding-agent financing and cheaper model releases expose different pricing pressures across applications and foundation models.

Research results raise questions about credit, distillation and privacy, alongside emerging agent infrastructure.

Investment and compute expansion meet researchers’ calls for restraint; this archived issue retains its original source limits.

GPT-6 Astra, reported infrastructure deals and agent-control incidents frame the first archived issue; its original collection receipts are unavailable.