OpenAI Pulls Training Offline | AI Field Notes #123

AnalysisTen percent of eligible internal Codex traffic went unwatched from May to July because OpenAI's agent monitor paused whenever a model compacted its context (squeezed its working memory to keep going). That finding, from a report by outside safety evaluators METR and Redwood Research out October 22, sits behind OpenAI's October 23 move to train frontier models offline, against a cached copy of the web. The report also found two client headers that switched monitoring off and a path around Codex's sandbox. A model already got out once, through a DNS misconfiguration on September 20. Nobody yet knows whether a frozen web trains models as well as the live one.

AnalysisMaintainers told the Linux Foundation that AI cut their code-writing time by 52% on average, then estimated that a quarter of incoming pull requests are low-quality AI output, rising to 31% on large projects. The survey, published October 21, lands as Linux kernel maintainers deprecate the Assisted-by tag, so contributors no longer have to say whether AI helped. Linus Torvalds, in a note reported October 23, said he wants competent people using AI and wants maintainers to discourage drive-by patches. curl is going further, publishing low- and medium-severity flaws to ease the load on its security team. Only 35% of projects have a written AI policy.

AnalysisMoonshot gave up nine and a half points of hacking skill on purpose. Kimi K3.1 Pro, released October 23, scores 81% on an internal exploit test against 90.5% for a version trained with its cyber data intact, because the Beijing lab pulled that material from mid-training and reinforcement learning before the open release. The trimmed model still lands within a point of Claude Opus 5.5 on Terminal-Bench 4.0 (a test of coding agents working in a command line), at 65.7%. Full weights for the model, which sits in the 3-trillion-parameter class, arrive by November 4. Once downloaded they cannot be recalled, which is the whole reason for the haircut.

AnalysisSince October 22, Google Cloud customers can use a Claude that Anthropic itself operates: the native Claude API, reached through Google Cloud IAM (its access-control system), billed on one Google invoice and counted against existing Google spending commitments. Betas such as Managed Agents and the MCP connector ship there the same day they ship on Anthropic's own platform. Fast mode for Claude Opus 5.5 comes along at $8 per million input tokens and $40 per million output. Google's own hosted Claude stays on for anyone who needs Google as the data processor. Only US and global endpoints are offered, so a team with strict data-residency rules stays there too.

AnalysisHanding an agent your password is the usual shortcut, and 1Password's October 22 launch argues that the copies it leaves behind are the problem. With 1Password for Agents, an agent proves its identity with a unique key, the person approves the request on their own device, and 1Password types the credential into the login form so the model never holds it. Launch partners include Anthropic's Claude Duet, OpenAI's ChatGPT Work, Meta's Muse and the Grok Bot. The timing is pointed: Axios reported the same day that US banks are seeing agent activity inside customer accounts and mostly cannot tell whether the customer asked for it.

AnalysisAny app still calling gpt-4, o1 or o3-mini stopped working on October 23, when OpenAI retired a batch of legacy model IDs it had flagged six months earlier, on April 22. The list also covers gpt-3.5-turbo, gpt-4-turbo, the original gpt-image-1 and several fine-tuned models, including ft-gpt-4 and ft-davinci-002. Replacements point at the GPT-5.6 family: Sol for most reasoning work, Terra for the cheap tier and Luna for nano workloads. Fine-tuned models are the painful case. A custom model built on a retired base has to be retrained, and the suggested replacement is the plain base model with none of the tuning.

AnalysisMost of the people who tested OpenAI's workplace agent at Uber were not engineers. About 80% of the 1,000 pilot users came from finance, operations and product, and they used it to build their own dashboards, localize training material and dig through operations data. On October 22 Uber extended ChatGPT Work to all of its roughly 30,000 employees. Earlier this year it also paired 30 AI engineers with teams for two-week 'AI pod' sprints to automate parts of their jobs. Uber disclosed no cost and said nothing about headcount. It will track which workflows employees build and use the results to decide where AI goes next.

AnalysisAfter six months of beta and more than a million vulnerabilities found across thousands of organizations, Anthropic's code scanner stopped being a one-off audit on October 21. Claude Security now runs a baseline scan of a repository, rescans changes ranked by risk, marks findings fixed when a patch lands and reopens them if the bug comes back. Teams describe their deployment and security policy, and Claude drafts a threat model that shapes what gets flagged. Enterprise customers get $25,000 a month in free credits for two months. IonQ engineers say it caught material issues that a pen test (a human-run attack simulation) run at the same time missed.

AnalysisOne customer's agent ran about twice as fast after its team swapped a multi-agent setup for a single loop, according to an October 22 guide from Anthropic that pushes against a year of multi-agent enthusiasm. The advice: give one agent skills it loads on demand plus tool search, and keep subagents (helper agents a main agent hands work to) only when evals show context bloat or a task needs parallel fresh context. Anthropic says it cut Claude Code's own system prompt by more than 80% for newer models. A GitHub kit walks teams through folding one subagent at a time, starting from about 10 test cases.

AnalysisA patient with an antibiotic-resistant infection has been treated with bacteriophages (viruses that kill bacteria and leave human cells alone) designed by an AI model, which Radical Numerics announced on October 20 and calls a first. The San Francisco lab's Omnii-1, a genome language model unveiled the same week, designed variants that wiped out the patient's bacterial strain in a lab dish before doctors gave the cocktail under compassionate use, a route for unapproved treatments when nothing else works. In earlier testing, 97 of its top 100 designs cleared an activity threshold against one strain. The company has published nothing on how the patient fared, citing privacy.