Postman rewrites coding agents without asking; GitBook breach; Revolut's 90% inference cut

🔥The day's one thing

🏢How organisations are changing

• Tano — is the ten-person AI-native startup in a winning Pragmatic Engineer essay by engineer Neil Kakkar, who went from 8–20 merged pull requests a month at PostHog to over 200, with human code review mostly replaced by a review agent that runs on every PR. Median time to merge is under an hour, and the revert rate is reported lower than PostHog's. This is the credible version of the 'reviews are the bottleneck' argument: real numbers, plus an honest list of new costs — context thrashing, lost shared context, and agent plan limits. Worth circulating to any leader still assuming review queues are permanent. One shipped bug had the agent parse currency minor units as floats and multiply by 100 — a classic money-handling mistake that the automated reviewer and no human caught. The Pragmatic Engineer

• Revolut — says it made self-hosted GLM models its default for coding work instead of commercial tools like Claude Code, and cut its AI inference costs by roughly 90 percent in under a year. The company runs the models on its own GPU cluster rather than paying per token to a frontier lab. A named, regulated company publishing a working exit from vendor-priced inference is the strongest datapoint yet that a large agent bill has a credible alternative at scale, with data control thrown in — the usual argument for staying put is now weaker. Discussion of the move cites a 56-GPU Nvidia B200 cluster at a blended 30 cents per million tokens, far below what coding-agent subscriptions cost per token. The Register (via Techmeme)

🧾Vendors, pricing and risk

• GitBook — says an attacker used AI agents to seize infrastructure credentials at 19:00 CET on November 2, and it responded by taking services offline, replacing every secret, and invalidating all customer API tokens. Data that may have been accessed includes Git Sync secrets, third-party API tokens, and questions asked to its AI assistant. If any team connects GitBook to your repositories or tools, rotate those tokens today — and note that commenters found the public status page still telling users they need do nothing. The CTO confirmed the incident in the thread and promised a full post-mortem, while commenters questioned what evidence actually distinguishes an AI-driven attack from a fast human one. Hacker News

  “the attack was a lot more developed than what we've been used to with humans, especially in terms of pace and adaptability”

• Codex — is at the centre of an unconfirmed but widely upvoted report that the setting allowing OpenAI to train on Codex conversations was switched back on without consent. OpenAI has not confirmed it, and a separate write-up found the company's own help-centre wording on the setting does not settle the question. If your people disabled training to keep prompts out of vendor pretraining, check the toggle today — and treat this as the argument for putting zero-data-retention terms in the contract rather than trusting a settings page. The Hacker News thread drew 413 points and 131 comments, with commenters reporting similar retention behaviour in Google's Gemini. Hacker News

• FlashProve — is free tooling from Anthropic that uses cryptography to show which model actually produced a reply, so a customer can tell when a routing layer quietly substituted a cheaper model for the one they chose and paid for. Model substitution behind gateways and routers is a genuine procurement risk once your agents sit behind intermediaries; this is the first mainstream answer, and adopting it costs nothing. Named contributors include researchers from EPFL, MIT and StarkWare alongside Anthropic engineers. Anthropic (via Techmeme)

• Falcon Guardian — Debuts an AI kill switch to stop internal AI agents that go rogue on corporate devices, alongside an integration with Claude Cowork that lets the security platform see what Anthropic's agents are doing. Agent containment is becoming a product you can buy rather than build; if agents already run on your endpoints, this is the control your security team will ask about in the next review. CNBC (via Techmeme)

📰News the board will ask about

• CrowdStrike — is buying Fiddler AI, which monitors what agents do inside enterprises, for around $400 million, and Dreadnode, which stress-tests AI systems for flaws before they go live, for about $95 million. Both fold into its Falcon platform as Falcon Observe and Falcon Guardian respectively. Endpoint vendors are bundling agent observability and testing into security stacks you may already own — expect agent security to arrive in renewals, not as a separate budget line. Fiddler had raised about $99 million; Dreadnode, founded in 2023, had raised $14 million. TechCrunch

• Anthropic — filed updated IPO numbers showing compute costs rose to 58 percent of quarterly revenue, from 48 percent the quarter before — a $10 billion compute bill against $17 billion of revenue, including $1.25 billion a month paid to SpaceX. The vendor most engineering organisations standardise on is structurally priced just above its input costs; expect pricing and packaging to keep shifting as those commitments bite. The same filing shows roughly $895 billion of committed spend, mostly compute, with the listing expected before Thanksgiving. TechCrunch

• Salesforce — signed a definitive agreement to acquire Cresta, whose voice agents handle customer calls end to end for brands including Marriott, Comcast and Alaska Airlines, in a deal reported at $2.7 billion, folding it into Service Cloud and Agentforce. If you run a contact centre or already buy Service Cloud, voice agents arrive inside a platform you own — and the deal highlights how weak 'call containment' is as a savings metric. Salesforce

• Cat Wu — The product lead for Claude Code and Cowork has left Anthropic, the second departure from the Claude Code team in two months after Thariq Shihipar announced his exit in October; she had joined in 2024. Not a crisis, but a fair question to put to your account team: two senior departures in two months on the product with the widest footprint in engineering teams is worth watching for continuity. Business Insider (via Techmeme)

🌍The big AI stories today

• ChatGPT — OpenAI's consumer product lead says search-style queries in the app now run at roughly a fifth of Google's volume, and weekly users grew by 200 million in two months to 1.2 billion — a distribution shift that reaches your customers and your support channels. Big Technology

• Tech giants' off-books AI debt nears $3.2 trillion. Nikkei tallies US tech giants' off-balance-sheet AI obligations at about $3.2 trillion, almost double in six months, with commitments at Google and Meta reaching 80 percent of assets — the financing that underwrites the clouds your contracts run on. Nikkei Asia

• ARTEX — CrowdStrike says China-linked operators pointed the open-source ARTEX agent framework at more than 200 organisations across 14 countries since October 17; at one Dutch institute the agent went from initial access to root in 12 minutes — a pace that resets what incident response has to assume. BleepingComputer

💡One takeaway