Apple's trade-secrets case against OpenAI expands to 11 more ex-employees
Apple's lawsuit against OpenAI, first filed 13 July over alleged trade-secret theft, has widened: court filings this week show Apple's investigation has identified 11 additional former Apple employees, beyond original defendants Chang Liu and Tang Tan, who may have witnessed or been involved in transferring confidential product information to OpenAI. One employee reportedly met with a defendant to discuss unannounced Apple products before an OpenAI interview; Apple is now seeking a preliminary injunction and expedited discovery. OpenAI says it "does not have, nor want" Apple's trade secrets, and blames Apple's own access controls for any residual data exposure.
Why it matters: This is the case escalating, not just continuing — a wider circle of implicated employees and a request for an injunction raises the stakes on how aggressively AI labs can recruit from hardware rivals.
SaferAI evaluation: China's GLM-5.2 refused zero offensive cyber or bio tasks, Claude Opus 4.7 refused all of them
A new SaferAI safety evaluation found that Z.ai's open-weight GLM-5.2 — which trails GPT-5.5 and Claude Opus 4.7 by only months on raw capability — refused none of the offensive cyber or biology tasks it was given during testing. Anthropic's Claude Opus 4.7 refused so consistently that evaluators couldn't even complete the CyberGym benchmark on it. Z.ai has published no safety framework, pre-deployment testing commitment, or risk assessment for the model.
Why it matters: The capability gap between frontier and open-weight models keeps closing; this is the first hard evidence the safety-infrastructure gap isn't closing with it.
Self-propagating npm worm ChainDrop hits 400+ packages, specifically targets Claude and VS Code configs
Microsoft Threat Intelligence uncovered ChainDrop, a self-propagating npm worm (a Shai-Hulud variant) that has compromised more than 400 packages across unrelated publishers by stealing npm publishing tokens and automatically republishing infected package versions. The payload specifically injects malicious code into Claude and VS Code configuration files (.claude/settings.json, .vscode/tasks.json and others), creating a second infection path that survives beyond the original npm install.
Why it matters: This is a supply-chain attack purpose-built to hijack AI coding assistants and editors rather than just steal credentials — a materially different threat model for anyone running agentic dev tools.
Nature Medicine: non-experts over-trust vague AI explanations, clinicians do best with none at all
A new MIT-affiliated Nature Medicine study tested explainable-AI approaches for skin-disease diagnosis and found non-expert users improved accuracy mainly through excessive deference to AI, trusting vague, confident-sounding explanations more than precise ones regardless of correctness. Primary care physicians performed best receiving only the AI's raw prediction with no explanation at all, since they weighed it against their own training rather than being persuaded by the rationale offered.
Why it matters: It's direct evidence that "explainable AI" can increase automation bias in non-expert users even as it appears to help — relevant anywhere AI tools reach frontline staff without clinical training.
MIT researchers unveiled a dual-arm robot that adapts its physical assistance during stroke rehabilitation exercises in real time, trained on healthy participants performing movements at varying effort levels and on therapists guiding the robot via force-sensing gloves to capture individual technique. Clinical trials are underway at a Munich rehabilitation center, aimed at the shortage of therapists available for the roughly 5 million people left with long-term impairment by the 15 million strokes that occur globally each year.
Why it matters: This is AI learning a specific hands-on human skill directly from practitioners rather than from a static dataset, a harder and different transfer problem than most AI-in-healthcare stories.
Federal government opens $3.5M grant to build a national AI resource center for Native American communities
The Administration for Native Americans posted a $3.5 million funding opportunity on 30 July to establish the AI3 Action Institute, a national resource center to help American Indian, Alaska Native, Native Hawaiian and Pacific Islander communities adopt AI while reducing the learning curve and strengthening tribal governance over its use. Applications close 27 August under opportunity number HHS-2026-ACF-ANA-NAI-0035.
Why it matters: It's a federal AI-adoption grant built explicitly around tribal self-determination rather than a generic nonprofit applicant pool — a funding lane most NP-AI coverage misses entirely.
Cooperative AI Foundation's $10M multi-agent safety research fund closes applications tonight
The $10 million research call from the Cooperative AI Foundation, Google DeepMind, Schmidt Sciences, ARIA and Google.org — first announced 17 June to fund research into risks from large populations of interacting AI agents — closes applications tonight at 11:59pm AoE (8 August). The Early-Career Track awards up to £100,000 (~$127,000 USD); the program funds original research only, not educational or advocacy projects.
Why it matters: Today is the actual action-deadline for a fund that was only covered here as an announcement back in June — worth a reminder now precisely because that earlier coverage doesn't tell anyone it's closing tonight.
Fortune op-ed: AI's new millionaires should fund existing nonprofits, not try to disrupt the sector
A Fortune op-ed published today by a tech founder and a nonprofit CEO argues the wave of new AI wealth — including roughly 4,400 new millionaires minted overnight by SpaceX's IPO this summer — should flow into existing nonprofits rather than fund attempts to rebuild the sector from scratch. They cite MacKenzie Scott's $26 billion in unrestricted gifts since 2019, where 90% of recipients reported stronger finances, expanded programs and more capacity to innovate, as evidence established organizations can absorb transformational money effectively.
Why it matters: It's a direct rebuttal, aimed at exactly this newly-wealthy AI-founder audience, to the "nonprofits move too slow" assumption that often pushes funders toward parallel structures instead of funding what already works.
AI Impact Hub