Mistral's trillion-parameter 'Le Chonk', Copilot CLI leaks secrets, and websites start locking agents out

Claude for Google Workspace — is a sidebar add-on that puts Claude inside Google Docs, Sheets and Slides, plus connectors that let it create and edit those files from a Claude conversation. It reads the document you have open, sees what you selected, and edits in place. Most of your clients' real work is not in a terminal; it is in a spreadsheet nobody wants to rebuild. This is the sort of thing that gets non-engineers past the demo stage. Two modes decide how much rope it gets: approval cards for every change, or accept-everything. On Enterprise plans the usual controls — compliance API, customer-managed keys, audit export — apply to the add-on too. Anthropic

Decisions API — is a new OpenAI endpoint in public beta that takes text or an image and returns one of three narrow answer types: a probability that something is true, a pick from a list you supply, or a score against levels you define. It is deliberately not a chat. A lot of agent plumbing is routing and triage, and paying frontier-model prices to classify a support ticket is one of the easiest cost lines to cut in a client's pilot. One model is supported, gpt-6-luna, at ten cents per million input tokens, and you are only charged for input — no output or cache charges. Images must be pasted in inline; hosted URLs are refused. OpenAI

Berth — is a free, open-source macOS app that starts Claude Code, Codex and similar agents on your own development machines rather than your laptop, then gives you a window onto what they are doing: folded command logs, edits as diffs, questions answered in line. Everyone running several agents at once hits the same wall — the laptop closes and the work stops. This is the unglamorous fix, and it does not require renting anyone's cloud. Two branches can be put in mirrored lanes side by side to compare their chats, diffs and terminals. It also ships nineteen ports of popular editor themes, which tells you something about who built it. Berth

WebMCP — lets a web application publish named actions — update-slide, send-email — that an AI agent can call directly, instead of the agent squinting at the screen and guessing which box to click. Builder.io has shipped a small add-on so coding agents can find and use those actions. Worth watching because it is the opposite bet to screen-driving agents: sites choosing to cooperate. If it spreads, 'can an agent use our product?' becomes a product decision for your clients rather than a scraping arms race. The spec says how a page registers tools for the browser's own agent but not how an outside agent reaches them — that gap is what the skill fills. Outside purpose-built apps, you need Chrome's WebMCP flag or origin trial, from Chrome 149. Builder.io

SAS AI Navigator — is a hosted register of every AI system running inside a company — agents, commercial models, open-source models — tied to the business use case each one serves and the policies and regulations that apply to it. It follows a deployment from pilot through production to retirement. Every governance conversation you have stalls on the same thing: nobody can list what is already live. Expect procurement teams to start asking for exactly this, and expect it to be a line item in your engagements. SAS is giving customers six months free, and partners Protiviti and Asenion supply ready-made policy packs with compliance questions and risk classifications for teams starting from nothing. Techstrong.ai

Cryptographic Context Injection — is a trick where a page serves an agent a scrambled blob plus a key; the agent unscrambles it using its own shell, then obeys the recovered text as though it had written it. Adversa AI used this to get GitHub Copilot CLI to pull a production environment file off disk and ship the contents to a server they controlled. This is the concrete version of the argument you make to clients about why agent permissions matter: the same instructions sent as plain text were refused as prompt injection, so no content filter was ever going to catch it. Start to finish took 28 seconds with nothing on screen indicating a file had left the machine, and it needed autopilot mode plus a lenient model — one Microsoft model ran it half the time while two GPT-5.6 models refused. GitHub validated the report but declined to call it a vulnerability. Adversa AI

Retailers and airlines start slamming the door on personal agents. Consumer agents that shop and book on your behalf are running into blocks. Amazon has deliberately cut off Meta's Muse; Walmart says its blocks were accidental, caused by a prove-you-are-human button the agent cannot complete; Delta says it has no arrangement letting an outside agent book its flights; Yelp admits only agents that have paid for data access. Users report trouble at eBay, Zillow, Adidas and others. When a client asks whether their site should let agents in, this is the live state of play — and the answer is turning into a commercial negotiation, not a technical one. Meta, Walmart, Stripe, Sierra, Genesys and others are now drafting an open standard for how agents identify themselves to businesses. Cloudflare's 15 September change to crawler defaults is suspected of catching agent traffic on ad-carrying pages. TechCrunch

SAP Joule Assistants — SAP used its Connect conference to move from strategy to shipped product: assistants for revenue recognition, financial disclosure, overhead accounting, trade compliance and supply chain, rolling out through early-adopter programmes into early 2027. It also announced it is buying Belgian skills-mapping company TechWolf and launched SAP Pay. This is what agentic delivery looks like at the back-office end, and SAP's pitch — that governance is inherited from the ERP rather than bolted on — is the argument your clients will hear from their incumbent vendor. A Governance Assistant arrives in Q4 and an access-and-security one in Q1 2027, bringing agents under the same identity controls as employees. SAP's line on back-office work is that there is no room for error. The New Stack

Project Glasswing — Anthropic has published what two large security teams got from early access to its Mythos models. The recurring finding is that ordinary scanners catch single bugs, while the model traces chains — several small weaknesses across different systems that only add up to something serious when combined. A useful counterweight to the week's breach stories: the same capability that makes agents a liability is being sold as the defence, and your security-conscious clients will have seen this write-up. Booz Allen says a single analyst covered eight production systems spread over 138 repositories in twelve days, work its leadership reckons would otherwise have taken a bigger team months. In one case the model followed a single configuration setting through two programs in two languages to prove a device's security key was never locked. Anthropic

Lambda — The GPU rental firm is taking in as much as $4 billion, valuing it at $14.5 billion before the money goes in, in what is likely the final private round ahead of a listing penciled in for 2027. Coatue and Blackstone are running it, and the company also took on a billion dollars of debt last week. Worth tracking because the compute bill behind every agent deployment is being financed by exactly these deals — and this one shows how concentrated the risk has become. The order book went from $15 billion in June to $50 billion in September, and most of that jump traces back to a single $35 billion commitment Anthropic made in late August. TechCrunch

Muse — Techdirt has pulled together the run of problems since Meta's consumer agent launched: a flaw that allowed spying on Mac users, a route to root access on the host machine by pretending to be a Muse agent, and reports of it reading private messages and uploading them despite instructions not to. This is the product your clients' staff are installing at home, and it is the reference case for why 'give the agent access to everything' is a bad default. Apple changed macOS privacy settings after a columnist found Muse citing an Apple Messages thread he never gave it permission to read. A Meta source told 404 Media that security teams were pushed to patch fast rather than delay the launch. Techdirt

PwC Global Workforce Hopes and Fears Survey 2026 — PwC polled 49,364 workers in 48 countries and found 64% had used AI at work in the past year, with daily use of generative tools climbing from 14% to 22% — while the share saying they get adequate job training fell from 59% to 51%. The number to quote when a client's adoption programme is all licences and no enablement; PwC's framing is a workforce splitting into speeds, with the biggest group left behind. PwC sorts workers into four groups, the largest being a 56% 'engine room' with little AI experience, where only two in five say they can get the training they need. Economic volatility worries more people (57%) than being replaced by AI (44%). Techstrong.ai