Relevance 10/10Importance 9/10
Anthropic opened a research preview of Claude Duet, an app built on Claude Opus 5.5 where the model works proactively — delivering a morning brief, running tasks in the background on its own cloud VM and browser, and checking in before acting on your behalf. It inherits memory, skills, and connectors from your Claude account, with reported access to 6,000+ connectors versus roughly 4,000 for OpenAI's Dots. Waitlist-only on Pro and Max in the US.
Relevance 10/10Importance 9/10
OpenAI and Google Cloud are putting GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna into Google Cloud's Model Garden, with Codex agentic workflows wired into Gemini's platform and Antigravity. Enterprises can buy through the Google Cloud Marketplace and draw down existing GCP spend commitments. Private preview for select customers now, GA "in the weeks ahead."
Relevance 9/10Importance 9/10
A study published Thursday in Science shows Microsoft's multi-agent diagnostic orchestrator correctly diagnosed hard NEJM-style cases 83% of the time versus 36% for primary care doctors and general internists working solo. On 325 emergency department cases it scored 80.3% against physicians' 46.8%. Big caveat: text-only historical cases, no specialists in the comparison, no real patients — a clinical study at Beth Israel Deaconess is planned.
Relevance 9/10Importance 9/10
Axios scooped that Morgan Stanley, Bank of America, JPMorgan Chase, Wells Fargo, and Citi have detected third-party AI agents logging into customer accounts using customer credentials — in some cases moving funds and calling support lines. Morgan Stanley has started blocking agents outright; others are restricting to read-only. Executives used the word "horrified," and flagged a systemic worry: agents could sweep deposits toward higher yields en masse.
Relevance 8/10Importance 10/10
Microsoft attributed the ConnectWise Cloud Backup compromise to Storm-0940, which used ARTEX — a Go-based agentic AI pentest framework — to automate recon, password spraying, and OAuth app abuse at machine speed. The agents hit more than 2,250 Microsoft 365 tenants at roughly 2,900 sign-in attempts per minute, with some intrusions reaching privileged access in under three hours. Targets span defense, energy, finance, semiconductors, and government.
Relevance 9/10Importance 8/10
Anthropic handed the Agent Skills standard — name, logo, GitHub org, and agentskills.io — to the Linux Foundation-housed Agentic AI Foundation, which also stewards MCP. Since publishing in December 2025, 70-plus platforms including ChatGPT, Cursor, GitHub Copilot, and VS Code have integrated it, with about a million skills published and nearly 280 million installs.
Relevance 9/10Importance 8/10
OpenAI opened its Decisions endpoint to everyone, running GPT-6 Luna against bounded-answer questions and returning structured outputs with probabilities. It cuts single-call latency from one-to-three seconds down to 100-180 milliseconds — up to 10x faster than Luna through the Responses API — at $0.10 per million input tokens. Box, Datadog, Ramp, and Speak were early testers.
Relevance 10/10Importance 7/10
Nick Cammarata — who co-named the superposition hypothesis with Chris Olah — now says superposition was the wrong explanation for neurons that appeared to mix unrelated concepts. His audits put roughly 90% of Qwen3-32B neuron poles as having one clear job, 7% with two related jobs, and only 2-4% unexplained; the famous "cat and car" neuron turns out to track pointed versus rounded shapes. If it holds, years of interpretability tooling gets simpler.
Relevance 8/10Importance 8/10
Research engineer Sofia Liguori and AI researcher Andreas Kirsch have left Google DeepMind, telling the BBC and The Intercept they lost faith in Google's governance of military uses of its models. Kirsch said oversight isn't "sufficient or meaningful enough to deploy AI safely and responsibly in military and classified settings." Several more staff are reportedly on the way out.
Relevance 8/10Importance 7/10
Canva Research trained Craft from scratch on Canva's own library of human-designed templates rather than photo-heavy datasets, running a planning pass for layout, hierarchy, and text before rendering up to 2K across 30 aspect ratios. It debuted fourth on Design Arena, ahead of entries from Google, Microsoft, Meta, xAI, and ByteDance. Live to all users in Dream Lab, plus API and MCP access from ChatGPT and Gemini.
