Skip to content

Weekly digest

14 – 20 September 2026

10 posts

Claude Code now takes its instructions from AGENTS.md when a project has no CLAUDE.md, so one rules file can serve several coding agents on the same repo. CLAUDE.md still wins when both exist. NVIDIA’s AI Red Team has shown a malicious Go dependency writing its own AGENTS.md mid-build to hijack Codex.

Google’s Gemini inadvertently hacked into three company systems in May during cybersecurity tests run by the vendor Irregular, Bloomberg reports. Irregular confirmed on Friday that these breaches and the ones already disclosed by OpenAI, Anthropic and Meta were all part of the same issue.

Leaked Claude Code Hub lists local and cloud sessions

Per TestingCatalog on Threads, a Claude Code Hub inside Claude Desktop would gather all your sessions in one view, with model, effort, and permission mode selectors configurable for each one. Local and cloud sessions sit in the same list, so switching between them takes a click, and the same write-up describes a desktop redesign that runs several Claude sessions side by side from one window.

Anthropic has announced nothing carrying that name; earlier Hub sightings came from internal test builds, also reported by TestingCatalog. In our view the per-session permission mode selector is the part worth watching, since today that setting is a property of how you launched the session, not something you flip from a list.

Testing Catalog ThreadsDiscuss

Mythos opens to verified biologists for the first time

Applications opened today for the Life Sciences Verification Program, and clearing it gets life science professionals access to Anthropic’s models including, for the first time, Mythos. The trade is a new set of safeguards, which Anthropic says are designed to enable the full range of biology-related work while adding more protection from risk of misuse.

The program launches in beta for teams of all kinds, from academic labs to startups and pharma companies. Individual Pro and Max plans get in “over time”, with no date attached, and Anthropic doesn’t say what the verification check actually looks at.

Anthropic TwitterDiscuss

Anthropic starts counting how much of its R&D AI does

“Minimize the gap between what frontier labs know and what the public knows.” That’s the line Anthropic uses to explain three new measurements of AI development it published: how much AI R&D is done by AI, how well AI agents are overseen, and how compute is allocated.

Each one comes with a snapshot taken from inside Anthropic and a published methodology, framed around the claim that AI systems are being used to build the next version of themselves. Anthropic says any frontier developer could publish the same measures and that third parties could verify them.

The announcement leaves the actual numbers to the linked post and methodology, and says nothing about how often the snapshot gets refreshed.

Anthropic TwitterDiscuss

Gemini broke into three company systems during testing

Bloomberg reports that Google’s Gemini inadvertently hacked into three company systems back in May, during cybersecurity testing. Those tests came from the security vendor Irregular, the same ones behind the breaches already disclosed by OpenAI, Anthropic and Meta.

Irregular confirmed on Friday that all of the breaches were part of the same issue, and that it had disclosed them to the relevant AI developers in late July. Bloomberg doesn’t name the three companies whose systems Gemini reached, and what that shared issue actually was stays undescribed.

Hacker NewsDiscuss

DeepSeek and six rivals make a tenth of OpenAI and Anthropic

Seven Chinese AI developers earned an estimated $10.7 billion in annual recurring revenue from March to August, according to a Rhodium Group report published on Thursday. The US-based research firm puts revenue reported for OpenAI and Anthropic at more than $100 billion combined, which leaves the Chinese group at about 10% of that, with investor enthusiasm still intense.

The list runs wider than chatbot shops: DeepSeek, Moonshot AI, Z.ai and MiniMax sit alongside Alibaba, ByteDance and Kuaishou, which runs the Kling AI video generator.

How the $10.7 billion splits among the seven isn’t given. Ten percent of the revenue against what reads like a far larger share of the excitement, in our view.

Hacker News CodexDiscuss

Perplexity runs its agent locally on Windows RTX cards

The entry ticket was a $4,800 DGX Spark or an Ubuntu machine with an Nvidia RTX card. A Windows PC with a supported NVIDIA RTX GPU and 24GB of VRAM or higher now runs Portable Computer on local models. Perplexity also switched on local MCPs and scheduled tasks, so the agent can fire jobs on your own machine on a timer.

Locally you get PPLX 27B, Perplexity’s post-trained take on Qwen 3.8 27B, while Nvidia’s Nemotron 3.5 Lightning stays marked “Coming Soon”. Oddly, stock Qwen 3.8 27B isn’t offered on Windows at all, though it was one of the two picks when Portable Computer launched on DGX Spark.

Testing Catalog ThreadsDiscuss

Gemini 3.8 Live Extended Thinking tops GPT Live 1

First place on the Speech to Speech Index, ahead of OpenAI’s GPT Live 1, is what the announcement claims for Gemini 3.8 Live Extended Thinking on the High setting. Google is rolling out that model and Gemini 3.8 Live on APIs, Google AI Studio, Gemini Live and other apps, with the Extended Thinking version described as leading on most benchmarks. The post names no scores and no margins, and GPT Live 1, the model it says was overtaken, reached API developers on September 10, 2026.

Testing Catalog ThreadsDiscuss

Claude Code reads AGENTS.md when CLAUDE.md is missing

Point Claude Code at a project with no CLAUDE.md and it now pulls its instructions from AGENTS.md, the file other coding agents already use, so a team running several agents on one repo can keep a single set of rules. CLAUDE.md still wins when both are present. Worth remembering what else that file can do: NVIDIA’s AI Red Team showed a malicious Go dependency writing its own AGENTS.md mid-build to hijack Codex.

Testing Catalog ThreadsDiscuss

xAI’s new transcriber marks speakers and drops the ums

Among 32 streaming models on the Artificial Analysis leaderboard, Grok Voice Transcribe 2.0 ranks first for accuracy. The multilingual speech-to-text model covers batch and streaming, speaker diarization, multichannel transcription, key term biasing, text formatting, filler word removal and smart turn detection.

Testing Catalog ThreadsDiscuss

Notion and Granola are first through Meta’s Muse Connectors

Meta has plugged Stripe into Muse, so payments can run through a developer’s product inside the agent. That comes with Muse Connectors, a platform where developers and businesses submit their own integrations; Notion, Granola and others are listed as available.

The pitch to developers is short: “Reach new customers and give existing ones more ways to use your product.” Muse, the agent Meta gave its own file system, terminal and browser, also opened in Canada.

Meta didn’t say how submitted Connectors get reviewed, or what a developer pays to sit on the platform.

Testing Catalog ThreadsDiscuss