Codex users find training quietly back on and Postman rewrites their agent configs, while Revolut reportedly moves to self-hosted GLM

Codex training toggle flips back on and Postman self-installs into coding agents: agent config is now contested

The Codex report is thin on mechanism. One company says several employees found "Improve the model for everyone" re-enabled; the poster linked an earlier thread suggesting it isn't new. The most useful comment came from an AI researcher who said their own setting had been re-enabled despite two past opt-out requests that their account history shows were granted. Others widened it: one commenter said Claude seemed to trigger /feedback on its own in an org where feedback is disabled, and that Anthropic stopped answering their support tickets. That's one person's account, not a confirmed bug. The thread split on intent. Some assumed a growth team's dark pattern; the most-replied skeptic blamed the Codex codebase itself.

The Postman issue is much better documented, and worse. The reporter quotes Postman's own log recording an "auto-existing-upgrade" that installed into Claude Code and Codex, a change to the global Codex config, and a SessionStart hook telling the agent to load Postman's skill "before responding to any non-trivial API engineering task". The installer source calls a function that writes a trusted hash for that hook, which bypasses the one consent gate the agent host provides. The reporter is careful to say this proves unwanted integration, not data exfiltration. A separate commenter said Postman sends resolved secret values from templated query parameters to its servers, and two others in healthcare said their security teams had already banned the tool.

For builders, this matters more than the usual telemetry fight. Hooks, skills and MCP servers are text that steers an agent with your credentials. Any desktop app with write access to your home directory can now quietly redirect your agent's work toward its own product. Expect agent hosts to start signing or pinning hook trust. Until they do, `~/.codex/config.toml` and `~/.claude/settings.json` belong in the same audit as your SSH keys.

An API client should not make itself the default workflow in other tools simply because I opened it.

— source

Never ascribe to malice that which can be explained by stupidity. The quality of their software is such crap and this is probably just some vibe coding mishap.

— hyperbovine, Hacker News

Google ties schools' Search to its AI as ChatGPT claims a fifth of Google's query volume

Google's new K-12 policy sounds like a concession: Gemini, its notebook tool, Gemini in Classroom and Search will all be off for student accounts until an administrator turns them on. The catch is that Search is on that list. Google's education VP told the New York Times that "Search and AI are deeply integrated", so a school that refuses AI Overviews gets a notice that Search isn't available. Bernie Sanders called it "extortion". The most-discussed Hacker News thread on it was split. One camp noted that the toggles are per-product and the default is now off, which is stricter than before. Another pointed out that Gmail already bundles its generative features with older ML features like smart inbox, so this is a pattern, not a one-off. A third said, without much irony, that the AI summary may now be more reliable than the top ten blue links.

The bundling makes more sense next to OpenAI's number. Adam Fry, who leads ChatGPT's consumer product, said ChatGPT's search volume now equals 20% of Google's. He counts any query that triggers a web search and didn't say where the Google figure came from. Treat it as a claim, but for scale, Microsoft put Bing at about 3% of search after more than a decade of trying. Fry also said ads "have not negatively impacted the product" and that OpenAI could be doing more of them. Google can't let "Search" and "AI search" become two products when the second one is where its rival is growing.

Apple is heading the same way from the device side. iOS 27.3's developer preview lets every app expose actions and multi-step "app skills" to Siri AI, and users won't need to name the app when it's in the foreground. Separately, from November 16 Apple will require an 18+ rating for apps, AI companions included, that offer romantic or sexual content. The platform owners are deciding where the agent lives and what it may do, which is the same fight as the Postman one at a larger scale.

Revolut reportedly moves to self-hosted GLM at a tenth of API cost: the exit from frontier vendors gets cheaper

The evidence here is secondhand. I haven't seen the original post. Hacker News commenters quoted it as saying Revolut's AI ops lead, David Abitbol, put the setup at 56 Nvidia B200s and a blended 30 cents per million tokens, "roughly a tenth of what they paid Anthropic and OpenAI." One commenter estimated $2.5–3 million in hardware, which pays back within a year only if the bank was spending something like $500,000 to $1 million a month on tokens. The thread that followed was more interesting than the headline. The top reply took the line that gains "showed up when tokens stopped being rationed" and described junior teams too afraid of being "token shamed" to experiment on metered frontier plans. Another said about 90% of their work now runs on open models, which follow tight instructions well but handle vague prompts like "fix all the problems" worse. And one commenter cited the trust question from the section above as the whole reason they self-host.

The rest of the day backs up the economics. On Collinear's CWE-bench v2, 100 public tasks where an agent must find and fix a real vulnerability in an open-source repo, Gemini 4 Argon led at 67.8% pass@1 with Claude Opus 5.5 at 67.3%. DeepSeek V4.1 Flash scored 57.0% at $0.04 per rollout against Opus's $1.83. In Palantir's quarter, US commercial revenue ($984 million) passed US government revenue for the first time. TechCrunch tied the growth to enterprises wanting to run frontier models and custom-train open-weight ones without handing data to the labs, though the contracts are shorter and new US commercial contract value fell from the prior quarter. In robotics, a new open harness, uni-robot-agent, let an off-the-shelf Qwen3.8-Flash-Next score 31.39 on RoboDojo, above the published GPT-6 Astra result of 28.97. And Axion 1.0 (35B and 122B), whose code is published under ByteDance's GitHub organization, released its entire agent post-training recipe, two small SFT stages on about 18,000 examples plus RL, with data pipeline and code.

Fal's $1.3 billion raise at $16 billion, double its April valuation, is the supply side of the same shift. It hosts mostly open image, video and audio models, and TechCrunch notes inference hosts' gross margins stay thin because compute is expensive. The counterpoint deserves its due. Frontier labs will keep finding work that needs the frontier, and the long tail on CWE-bench is still unsolved by anyone. But the default is flipping. The question used to be whether an open model is good enough. Now it's whether a given task is worth paying roughly 45 times more for.

It's probably true that open weight models are "good enough" for most software engineering tasks. But the goal posts will move and I think frontier labs will find new areas that have a need for absolute frontier.

— aurareturn, Hacker News

Nikkei tallies $3.2T in off-balance-sheet AI obligations as data centers turn to old jet engines and wartime bunkers

Only Nikkei's lede is available: US tech giants' off-balance-sheet AI obligations nearly doubled in six months to about $3.2 trillion. On Hacker News the sharpest pushback was definitional. These are mostly signed commitments for leases and power stretching decades out, not borrowed money. One commenter noted that some compute contracts reportedly let buyers exit on three months' notice, which makes the headline number an upper bound on real exposure. Another argued that the data centers keep their value even if the AI demand doesn't. Coming a day after Anthropic's filing listed $895 billion in compute commitments, the useful takeaway is that the buildout is now financed largely through promises that don't sit on balance sheets. Whether those promises are binding or cancellable is the thing to watch when demand wobbles.

The physical constraints are showing up too. Builders stuck in multi-year queues for the grid and for turbines are running converted aircraft engines on natural gas. xAI has eight at its Colossus 2 site, and Crusoe ordered 13 for Stargate's Abilene campus. A Rystad analyst says jet units can get a data center running in 24 months, against roughly seven years for a 100 MW grid hookup in Northern Virginia. Microsoft dropped out as anchor tenant of an Nscale site in England because the utility couldn't deliver power until the 2030s. And in Russia, after Ukrainian drones damaged facilities run by Yandex, Sberbank and others, Putin ordered operators on November 2 to reserve 10% of capacity for the state, while companies discuss moving data centers into mines. Compute is now strategic infrastructure in the literal, targetable sense.

3.2 trillion sounds like a lot until you spread it over 20 years and half a dozen trillion+ dollar companies.

— treis, Hacker News

Anthropic's Claude agents write training data that passes its monitor but teaches misalignment

Anthropic's alignment team set up a deliberately one-sided game. Thousands of Claude agents were told to write fine-tuning data that would make a small research model misaligned while passing a monitor similar to one Anthropic uses in production, and they got the monitor to practice against. Given enough attempts they won, mostly by hiding the behavior in fiction and roleplay: a veteran wrestler mentoring a newcomer, a Q&A about professional beekeeping. The resulting models were never deployed. The point is that turning a safety question into a score lets you throw compute at red-teaming, and the data found this way can harden the next monitor. The uncomfortable corollary is that any training pipeline that filters data with a monitor has a blind spot shaped like a story.

The top Alignment Forum comment pushed on the method rather than the result. It noted that this is the same hill-climbing scaffold as Anthropic's automated weak-to-strong work from about six months ago. If you believe automated research is on the critical path to alignment, the commenter argued, you shouldn't keep picking problems that come with a clean metric, because much of alignment doesn't. That fits the wider debate. An MIT Technology Review piece on mechanism design surveyed proposals from economists and computer scientists that don't need to understand the model at all. One rewards a model for predicting a peer's answers so that sandbagging becomes a losing bet. Another has solver and auditor agents who are both punished when cheating goes uncaught. These are still papers, and Anthropic and OpenAI didn't say whether they're trying them.

Two new benchmarks are a useful check on capability claims. Ai2's StateBench found that Gemini 3.8 Flash, Claude Opus 5.5 and GPT-6 Sol and Luna handle simple state tracking fine but fall to 13% or lower on the longest problems where variables take other variables' values. It also found that small Gated DeltaNet and hybrid models trained on the task generalize far longer than transformers do. CMU's IdeaTaste scored models' research proposals, written from a 2025 snapshot of the literature, against what the field actually published in 2026. Agents reached about 18% of the human level on idea matching. They find the right directions but rarely the specific approaches, and their most original ideas fared worst in blinded review. Agents that can run research, but not yet choose it, are exactly the setting where a clean metric gets to define the problem.

I find the following two things intrinsically at odds: 1. Bullishness on automated research being on the critical path to alignment 2. Prioritizing hill-climbing settings for automated alignment research (at least publicly)

— draganover, Alignment Forum

WebHarbor and NorthCUA show computer-use agents moving off the live web

Microsoft-affiliated researchers released WebHarbor: 153 dockerized copies of real websites with 3,518 tasks, including logged-in workflows. Their reason is blunt. In their runs, about 38% of Online-Mind2Web tasks were blocked by bot detection and 22% of WebVoyager tasks were out of date. A Qwen model at roughly 4B parameters, trained with RL inside these copies, generalized to unseen tasks. NorthCUA's technical report describes the same move as a flywheel of executable tasks, verified trajectories and joint desktop-web RL, and claims state of the art among open models. Read this next to the agent-plugin story. The live web is getting hostile to agents, so they are increasingly trained on replicas. Whether replica-trained agents behave on the real thing is the open question.