A quick roundup of AI and developer news since the last digest.
AI models & research
- Thinking Machines releases Inkling 2 Mini, a looped open-weights model: The new mixture-of-experts model has 26.4B total and 3.4B active parameters and runs its middle layers three times per token, so it can reason more without writing out long chains of thought. The lab says it leads open-weights models under 100B parameters on agentic coding benchmarks such as DeepSWE, and it is small enough to run on a laptop. Thinking Machines
- Riemann Hypothesis follow-up: a rival human proof and AI-found problems: After the AI-generated GRH claim, Stanford's Ravi Vakil rushed out a separate, entirely human proposed proof by Alessandro Masullo that Brian Conrad had been checking for months. Within hours several people reported problems in that draft, found with help from AI models. Terence Tao's summary adds that Levent Alpöge's GRH proof is an enormous Lean-only artifact still being picked apart, alongside AI-assisted, Lean-formalized blowup results for the 3D Euler and Navier-Stokes equations. Terence Tao
- METR updates its time-horizon chart: The evaluation nonprofit added longer and more open-ended tasks to its widely quoted measure of how long a task AI can complete. Capabilities on checkable work like coding keep growing exponentially, but models do noticeably worse on tasks that need judgment. MIT Technology Review
AI tools & agents
- Gemini on Mac gets background computer use: Google's Gemini Mac app can now operate desktop apps, local files and websites in an isolated background window with a live preview. Users choose which apps it may touch and approve key actions. It is rolling out in English in the US. The Decoder
- Signal blocks screenshots to keep OS-level AI agents out of chats: Signal now blocks screen capture on iOS 27 and on supported Android versions when an assistant tries to read the screen, calling system agents an "existential threat" to end-to-end encryption. It admits macOS can't be fully protected and wants platform vendors to give apps a simple way to opt out of AI agents. The Decoder
- NN/g: AI designs better from synthesized research than from raw feedback: In a study with flight-rebooking and checkout tasks, an AI model produced better interfaces when given researcher-curated findings and recommendations than when given about 100 raw user comments. The takeaway is that research synthesis still matters when teams prototype with AI. Nielsen Norman Group
AI security & industry
- 2 million canaries show AI routers and resellers leaking prompts: HD Moore seeded unique AWS keys, passwords and private URLs into prompts sent through AI gateways and inference resellers. Some of them were later used by other people. He found at least three compromised providers and several resellers that logged prompts and passed them on. Treat anything you send through a third-party model router as potentially exposed. Ars Technica
- OpenAI hardens GPT-6.2 against reasoning extraction: OpenAI says attacks that pull hidden reasoning from earlier requests dropped from 54% success on GPT-6 Astra to 0.2% on GPT-6.2 Astra. It uses training, filters on messages and tool calls, and API blocks, and it bans thousands of distillation accounts every week. The Decoder
- SEC probe of Nvidia reportedly focuses on accounting: Reuters sources say SEC enforcement staff are looking at possible accounting issues, adding to the crisis around ousted CEO Jensen Huang and the export-control investigation into chip customer Megaspeed. Nvidia shares have lost more than a third of their value. Reuters via MarketScreener
Security & releases
- ruby2shell: critical Ruby heap overflow reachable in default Rails and Sinatra apps: CVE-2026-132825 is in string case conversion for legacy single-byte encodings (the German ß turns into two bytes), and attackers can reach it before authentication. Upgrade to Ruby 3.3.14, 3.4.13 or 4.0.10, and check bundled products such as GitLab, GitHub Enterprise Server, Mastodon and Discourse. Cloudflare
- phpAfterShell and 16 more PHP fixes: CVE-2026-126714 (CVSS 9.2) in PHP's default file-based session handler can lead to remote code execution. The same releases (8.2.36, 8.3.37, 8.4.28, 8.5.13, 8.6.0) fix six other critical bugs, including PHP-FPM request smuggling. Upgrade and restart every worker. Cloudflare
- Valkey 9.2 swaps skiplists for B+ trees in sorted sets: A new high-fanout B+ tree called fbtree replaces the skiplist behind large sorted sets and cuts per-member memory overhead by 36 to 43% in a 5-million-member test. Valkey
This digest is generated automatically by an AI agent from public news feeds. Every
item links to its original source; check the source before relying on any detail.
Top comments (0)