Morning Edition · Wednesday, October 7, 2026
Tech
Microsoft and Meta cut internal Claude spending weeks before Anthropic's expected listing
Reports say Microsoft lowered its projected internal Claude spend by more than a third and Meta's Claude Code user count fell by about half, as both companies move engineers to their own and partner models.
Tech
OpenAI publishes 722 math manuscripts from an unreleased model, with Lean proofs for part of the set
The release covers 372 result families drawn from about 4,000 attempted problems, and its own repository warns that results without formal proofs could contain errors.
Tech
Wikimedia links OpenAI agents to unauthorized wiki edits and attempts to use its tools as proxies
The foundation says agents traced to OpenAI infrastructure sent millions of API requests and hundreds of thousands of Wikidata queries, which may have contributed to a May outage.
Tech
No frontier lab tells New York City Council it carries catastrophic-risk insurance
Representatives of OpenAI, Anthropic, Google and Meta did not raise their hands when Speaker Julie Menin asked about insurance, as the council weighs a package of ten AI bills.
Tech
OpenAI will watermark ChatGPT and Codex text in the European Union to comply with the AI Act
The statistical watermark, called textGrain, was detected in about 95% of 400-token passages in OpenAI's tests, but replacing one word in ten with a synonym cut detection to about 66%.
Tech
Google releases EmbeddingGemma 2, a 740 million parameter open model that embeds text, code, images, video and audio
Modular encoders let developers load only the text-and-code block, which Google says runs in about 191 megabytes of memory on a Pixel 11 Pro, and the context window grows to 8,000 tokens.
Tech
AWS and Ai2 ship small open specialist models built on Alibaba's Qwen base weights
Strands Decider 2B replaces Qwen3.5-2B's language head with a pointer head that scores options in one forward pass, and AstaBrief 8B fine-tunes Qwen3-8B to write cited research reports 3.5 times faster.
Tech
Three new papers move prompt-injection defense for agents to the point of action
APEX enforces checks where agents call tools, MCP servers and skills, DIBench tests mobile GUI agents against deceptive screens, and Proxy Confidence uses a surrogate model's log-probabilities to audit closed agents.
Tech
New papers question what reinforcement learning guidance, backdoor leaderboards and summary-tree retrieval actually measure
One paper gives a bias-variance theory of when external guidance helps reinforcement learning with verifiable rewards, another says backdoor-defense costs require a never-poisoned control model, and a third retests RAPTOR-style trees at matched cost.
Tech
OpenAI's GPT-6 Astra scores 55% on Ironclad contracting tasks as Atlassian expands its OpenAI deal
OpenAI reports GPT-6 Astra beat GPT-5.6 Sol's 41.6% on 11 real contracting tasks and nearly halved time per attempt, while Atlassian commits more spend but keeps its platform multi-model.
Tech
Meta's open vision models move into national-lab science and assistive robotics
Berkeley Lab says Segment Anything Model 3 and DINOv3 cut beamline image analysis from a month of expert labeling to 15 minutes, and a University of Pittsburgh robot uses the same models on battery-powered hardware.
World
Giorgia Meloni files her voice as a European Union trademark to block AI clones
The Italian prime minister's application at the European Union Intellectual Property Office covers a four-second recording of "Io sono Giorgia Meloni," and a grant would apply in all 27 member states.