OpenAI's agents caught probing a UN site, a $78k Codex bill fight, and the AI roll-up pitch going viral

GitHub Spec Kit: structured specs for coding agents, and the community is all in. Spec Kit is an open-source toolkit from GitHub that gives a coding agent a defined process to follow: write down what you're building as a specification, plan against it, then let the agent execute with a documented trail — plus extensions for assessing ideas and fixing bugs. It's the clearest sign yet that spec-driven development — telling agents what to build in a reviewable document instead of a loose prompt — is going mainstream. For clients, that means prompt-craft becomes governed engineering artefacts their teams can audit and approve. The repository has racked up around 139,000 stars, and the bug-fixing extension installs with a single CLI command. GitHub

The enterprise coding-agent contract comparison you'll want open in your next procurement meeting. An article puts GitHub Copilot, AWS Kiro, Cursor, Devin and Windsurf side by side on the terms that actually decide enterprise deals: who pays if generated code infringes someone's copyright, where your prompts and data live, and what a big deployment costs — with vendor contract language checked on September 26. No vendor publishes these side by side, and the differences are stark enough to change a recommendation. This is exactly the homework a client's legal and procurement teams will expect you to have done. Copilot's Business and Enterprise tiers offer uncapped IP indemnification for unmodified outputs, while pinning inference to US or EU data residency adds 10% to AI credit consumption; Cognition's contract defines outputs as customer data, excludes them from indemnity, and caps exposure at twice the trailing twelve-month fees. MarkTechPost

Australia's parliament wants Altman and Amodei under oath. The two chief executives have been asked to appear before an Australian Senate committee examining AI and data centres — the direct political fallout from reports that OpenAI's browsing agents touched government websites in Australia and the US. It's the first time the agent incidents have produced a parliamentary summons, and a preview of the accountability format every jurisdiction is rehearsing. The committee, chaired by Senator Sarah Hanson-Young, sits again in Canberra from October 1, and a federal environment minister has already publicly described OpenAI's conduct as flatly unacceptable. explainx.ai

Amodei heads to the White House for a first one-on-one dinner with Trump. Anthropic's CEO will dine with the president at the White House — their first meeting alone together, at a moment when the two sides disagree openly about how fast AI should be built. Anthropic is the safety-first lab, and it's currently fighting a Pentagon supply-chain-risk designation in court that followed its attempts to restrict how its models are used. This dinner is a data point in whether the US government's posture toward AI safety is hardening or thawing. The Pentagon designated Anthropic a supply-chain risk earlier this year after the company put guardrails around use of its technology; Amodei was also lampooned on Saturday Night Live's season premiere. TechCrunch

Unsealed filings: OpenAI's own executives knew the book piracy was a problem. Newly public documents in the Authors Guild class action against OpenAI and Microsoft allege that executives at both companies knew their models were trained on pirated books — including Microsoft being told about the LibGen library as early as April 2019, when GPT-3 was presented to Bill Gates. This case has been running for years, but the internal language coming out now is unusually candid — and its outcome will shape what training-data liability looks like for every AI vendor your clients buy from. The filings say OpenAI deleted its LibGen files in the summer of 2022 in an effort dubbed 'Project Clear', and that a researcher hired to improve model writing was given finishing George R. R. Martin's series as a stated mission. The Authors Guild

“a bunch of artists express worry about what we're doing here and we'll likely ignore their concerns and release anyway”

Google lets Gemini users buy from Flipkart — agentic shopping gets a checkout. In a live test in India, some Gemini and AI Mode users see a Buy button on Flipkart product listings that opens the retailer's checkout without leaving the AI interface. After last week's story about retailers blocking shopping agents, this is the opposite bet: a retailer and an AI company cooperating on in-chat purchasing. Expect clients to ask which pattern wins. The test covers a limited set of products — smartphones, electronics and mobile accessories — with a broader rollout planned for October, ahead of India's festive shopping season; Google put about $350 million into Walmart-owned Flipkart in 2024. TechCrunch

TechCrunch asks whether anyone will trust Meta's consumer agent with their money. Meta unveiled Muse, its ground-up consumer AI agent, at its Connect event — and the immediate conversation isn't about what it can do but whether people will hand a company funded by advertising their sensitive personal and financial details. It's a clean illustration of the market splitting: Meta is betting on consumers while OpenAI and Anthropic head toward enterprise and public offerings. Which side of that split a client sits on determines which trust problems they inherit. A podcast host who tried Muse says it found him unclaimed money but calls it a one-time party trick rather than a habit — and notes it didn't plug straight into his Threads, Instagram and Facebook data, which paradoxically made him more willing to use it. TechCrunch

“Meta's business is to sell you ads.”