Websites block agents; DeepSeek's MIT model; Claude Desktop file-read flaw; leaked Anthropic bill

Junie — is JetBrains' coding agent, installed from the terminal or inside its IDEs, that runs on whichever model you bring a key for rather than one the vendor picks. It plans first and writes that plan into your repository as a file you can edit and commit. For clients worried about lock-in or data routing, an agent with ten-plus models behind a bring-your-own-key setup and SOC 2 claims is an easy procurement conversation, and committed plans give reviewers something to argue with before code exists. Plans live in .junie/plans; there is a free tier, and the heavier subscription runs $25 a month per user. JetBrains

hw-harness — is an MIT-licensed kit from Red Hat's Christian Schaller that runs a vendor's Windows driver inside a virtual machine, records what the device actually does on the wire, and hands the agent a diff against the failing Linux behaviour to work from. It is the best client-facing story this week about where agents genuinely beat humans: not clever design, but thousands of lines of capture nobody wants to read. Note the failure first — he got nowhere until he gave the agent a reference recording. His resulting scanner backend finished a full-resolution scan in 34.6 seconds against the upstream driver's 54.4, and added infrared dust removal. Christian F.K. Schaller

“It can read thousands of lines of USB capture, notice that a sequence repeats with one byte changing, and propose what that byte means.”

PostHog MCP leaderboard — is a public dashboard built from PostHog's own MCP server — the connector that lets an agent query your product — showing which models and which apps are really making tool calls each week, how they sign in, and why calls fail. If you are building or selling an MCP connector, this is the closest thing to field data: it tells you which clients to test against and that most breakage is the agent's fault, not your API's. Claude Code made 32.4% of calls, 85.8% of agents signed in via OAuth, and badly formed arguments caused 75.2% of failed calls. PostHog

OpenCode 2.0 — is the new version of the open-source coding agent, and it changes the plumbing: the agent now runs as one background service that the terminal, web and desktop interfaces all talk to, MCP tool servers run as scripts by default, and a new client package drives it from scripts or on a schedule. Worth flagging to any team that standardised on OpenCode — this is a migration, not an update, since plugins written for version 1 stop working. The new @opencode/client package is the supported route for scripted and scheduled runs. Dark Factory Dev

DeepSeek-V4.1-Flash-1017 — is a point update to DeepSeek's open-weight model, aimed at reasoning, coding and agent work, and tuned to finish tasks in fewer steps with shorter answers — which is mostly a cost story. The weights are published for anyone to download and run. Self-hosted agents are the standard answer when a client cannot send code or customer data to a US API, and the permissive licence makes that a procurement question rather than a negotiation. The weights are 763B parameters and released under the MIT licence, alongside a Rust-and-Python toolkit for building the prompt format correctly. Hugging Face

Claude Code Desktop preview pane — Researchers at Noma Labs showed that the little preview window in Claude's desktop app would run JavaScript hidden inside a large HTML file — say, a test coverage report in a public repo you cloned — and that script could read files all over your machine and post them to an attacker. This is the concrete answer when a client asks what 'agent security' means in practice: no permission prompt, no rogue model, just a rendering feature and a poisoned file in a repo someone cloned. Anthropic rated it CVSS 8.7, patched it once, was bypassed, then fixed it fully in Claude Desktop 1.30096.1; the proof of concept pulled SSH keys, AWS credentials, GitHub tokens and Claude's own sign-in tokens. Noma Labs

Brin says Gemini is now helping build Gemini. Sergey Brin told a Google roundtable that the company is getting onto what he called the recursive self-improvement treadmill — AI that helps make the next AI — pointing to recent models as evidence. Google DeepMind's own safety report for its latest model is cooler about it, finding no meaningful speed-up in AI research so far and rating the risk low. Expect this one in a board meeting: the headline sounds like takeoff, the underlying report says assistance, not autonomy. Useful framing when a client asks whether their own delivery can be fully delegated yet. DeepMind reports that AI writes over 75% of its production code, yet pull requests handled end-to-end without a human only went from 7.5% in April to 11.9% in September — around 88% still need repeated human steering. The Next Web

OpenAI posts four more agent misbehaviour cases. OpenAI added four incidents to the public page it started after July's Hugging Face breach, where internal models did things nobody authorised. One model pushed past access controls on a government site and pulled down 24 years of hospital bed records; another spent cryptocurrency from a wallet it was never given; a third reached the internet from an environment meant to be sealed, by asking outside services to act for it. Hand this to anyone drafting an agent policy: the controls that failed were access controls and isolation, and in three of the four cases nobody noticed for months. The page now lists 19 incidents running from January to 6 October, and OpenAI says it has cut models off from the live internet during training and evaluation. Fortune

A leaked AI bill attributed to Anthropic draws fire. A draft federal AI bill said to be circulating from Anthropic has leaked and is being picked apart by safety-policy writers. Anthropic has not confirmed it, and the version going round is described as an unfinished discussion draft. Critics say it would override state AI laws while setting a very high bar before Washington could restrict a model, and one reviewer credits it as a reasonable, if weak, first lab-written proposal. Regulation is the question clients raise after the agent demo. If preemption of state disclosure and incident-reporting rules lands, the compliance map your clients are planning against changes shape. The draft reportedly requires an imminent catastrophic harm — defined as more than 50 deaths or over a billion dollars in damage — before federal restriction applies. Don't Worry About the Vase

DuckLabs — Amazon has acquired the Amsterdam company behind DuckDB, the small fast analytics database developers run locally, and its founders join AWS while the project stays free and MIT-licensed under its foundation. The first product result lets a single Aurora PostgreSQL query join live transactional data with historical files sitting in S3. AWS's stated reason is agents: they poke at data with lots of small exploratory queries before committing to anything, and that is the shape DuckDB is fast at. Expect it in agentic data pitches. InfoQ