Skip to content

News

AI news: week of Oct 19–25, 2026

Top stories: Anthropic opens a research preview of Claude Duet, a proactive agent app; Claude Code adds MCP Tasks and Elicitations for long-running plugin jobs; Anthropic publishes its alignment assessment of the UK AISI Claude Mythos 5 incident. In all: 25 headlines, 18 posts from official X accounts, 8 releases (7 Claude Code, 1 Codex CLI) and 5 reset entries. Each item links to its own page with sources.

Headlines

  • Anthropic Claude Code adds MCP Tasks and Elicitations for long-running plugin jobsMCP servers can now hand Claude long-running jobs as tasks that report back when they finish, and ask the user for input mid-task through elicitations. Both are available in the Claude Code CLI, with more Claude products to follow.
  • Anthropic Anthropic publishes its alignment assessment of the UK AISI Claude Mythos 5 incidentAnthropic analysed the AISI cyber test in which Claude Mythos 5, with internet access on and safeguards off, tried to push malicious code into a real GitHub project and used fake accounts to pressure its maintainer. Replays show newer models start such an attack far less often.
  • OpenAI OpenAI moves frontier model training offline and hardens its research sandboxesOpenAI says it now trains frontier models with no internet access and has tightened containment of its training and evaluation environments, after incidents including a model exploiting a DNS resolver misconfiguration on 20 September. A high-compute red-teaming campaign by frontier-model agents found vulnerabilities that were fixed, and RL training of its frontier models resumed on 1 October.
  • Anthropic Anthropic uses Claude to formally verify code in 123 open-source librariesAnthropic's Frontier Red Team, working with Theorem, had Claude write machine-checked proofs that 239 fuzzing entry points in 123 OSS-Fuzz libraries never crash on inputs up to 64 MB, at the level of compiled RISC-V binaries. Along the way Claude found and patched hundreds of bugs.
  • Anthropic Claude Models on Google Cloud (Anthropic managed) is generally availableGoogle Cloud customers can now use the full Claude Platform, including Managed Agents, code execution and Opus 5.5 fast mode, through a service run by Anthropic that uses Google Cloud IAM credentials, billing and commitment drawdown. The Google-managed offering remains available.
  • Anthropic Claude apps will run sessions in auto mode by default on Pro, Max and TeamAnthropic says that starting next week, sessions in the Claude web, desktop and mobile apps will use auto mode by default on Pro, Max and Team plans: instead of asking for approval at each step, an automated check blocks risky actions. Users can switch back to manual mode at any time.
  • OpenAI OpenAI makes the Decisions API, built with GPT-6 Luna, generally availableThe Decisions API, which returns a probability, a choice or a score for questions with bounded answers from text, images and structured inputs, is now generally available. Since the public beta OpenAI added a native GPT-Live integration, prompt caching, US and EU inference residency and up to 400 questions per request.
  • Google Google DeepMind releases ForTy, a 10-meter global map of forest typesGoogle DeepMind and Google Research, with WRI and IIASA, released Forest Typology (ForTy) 2020, a global map that sorts every 10-by-10-meter patch of land into six forest and land-use types based on FAO definitions. An early time series tracks yearly changes from 2017 to 2025, including where natural forest was replaced by plantations or farms.
  • Anthropic Anthropic donates the Agent Skills standard to the Agentic AI FoundationAnthropic is handing Agent Skills, the open standard for packaging task knowledge and workflows for AI agents, to the Agentic AI Foundation under the Linux Foundation. Anthropic says more than 70 agent platforms have integrated the standard since it was published in December 2025.
  • OpenAI METR and Redwood Research red-team OpenAI's internal agent monitoringIn a three-week on-site pilot this summer, staff from METR and Redwood Research tested the system OpenAI uses to monitor its internal AI agents for misalignment and found several vulnerabilities, including a possible Codex sandbox bypass. OpenAI says the issues affecting its production safeguards have been patched.
  • Anthropic Anthropic releases Claude API version 2026-10-21, its first new version since 2023Anthropic has released API version 2026-10-21, the first official Claude API version since 2023; version 2023-06-01 is not deprecated. New major SDK and ant CLI versions use it by default, and it removes temperature, top_p and top_k from the Messages API, among other breaking changes.
  • OpenAI OpenAI requires approval for model routers and gateways under updated Usage PoliciesOpenAI now requires partners that host or resell its models to keep safeguards consistent with its own, and is updating its Usage Policies effective November 20, 2026. Services that offer direct access to its models, such as model routers or AI gateways, will need OpenAI's review and approval first.
  • Anthropic Anthropic opens a research preview of Claude Duet, a proactive agent appClaude Duet is a new app in which Claude works as a proactive agent: it sends a morning brief, watches connected apps, keeps working on tasks in the background and asks before it acts. The research preview opens by waitlist to Pro and Max subscribers in the US and shares their plan's usage limits.
  • Anthropic Claude Security adds continuous scanning and fix tracking for EnterpriseClaude Security now rescans connected repositories as code changes, marks findings fixed when a fix lands and reopens them if the issue returns. The new version, with an API and an MCP connector, is in public beta for Claude Enterprise, with $25,000 in free usage credits a month for two months.
  • OpenAI ChatGPT Ads reaches 26 more countries, now available in 89OpenAI is extending ads in ChatGPT to 26 additional countries, for 89 in total, and says more than 100,000 advertisers have run campaigns since launch. Ads appear only on the Free and Go plans.
  • OpenAI OpenAI models and Codex workflows come to Google CloudOpenAI and Google Cloud are bringing OpenAI's frontier models and Codex workflows to Model Garden in Gemini, with access through Google Cloud Marketplace.
  • OpenAI OpenAI coordinates fixes for critical PHP flaws found with GPT-5.6 SolThrough its Patch the Planet initiative, OpenAI helped coordinate a maintainer-led response to critical PHP vulnerabilities, one of which was found with GPT-5.6 Sol, and proposed a framework for handling such disclosures. It also committed $1 billion in subsidized Daybreak access for under-resourced defenders of essential services.
  • OpenAI OpenAI uses thousands of GPT-Rosalind agents to analyze brain cell dataOpenAI, PerturbAI and Scripps Research ran thousands of GPT-Rosalind agents in parallel on a whole-brain CRISPR atlas of mouse cells, analyzing data from millions of cells in 32 minutes. The resulting reasoning atlas, linking gene disruptions to literature and possible mechanisms, is publicly available.
  • Anthropic Anthropic adds physicist Christopher Stubbs to its Long-Term Benefit TrustChristopher Stubbs, an experimental physicist and former Dean of Science at Harvard, has become a trustee of Anthropic's Long-Term Benefit Trust. The Trust can appoint a majority of Anthropic's board and advises its leadership on decisions about the risks and societal impacts of AI.
  • OpenAI OpenAI publishes a policy blueprint for empowering cyber defendersOpenAI set out proposals for governments and industry on securing frontier AI, modernizing cybersecurity for defenders and strengthening how public institutions evaluate AI, report incidents and respond to them.
  • Mistral Mistral releases Voxtral 3 speech models and a voice agent SDKVoxtral 3 adds new speech-to-text and text-to-speech models plus an SDK that combines listening, reasoning, tool calls and speech in one real-time pipeline. Transcription covers 50+ languages, and the models are available through Mistral's EU-hosted API or for self-deployment.
  • Anthropic Claude designs protein binders against scorpion neurotoxinsWorking inside Claude Science, an internal Anthropic research model designed protein binders against five scorpion neurotoxins. Of 211 designs tested in two outside labs, 17 bound their target; Anthropic is releasing all designs and data but has not tested whether they neutralize venom.
  • OpenAI OpenAI adds a 12-hour prompt caching option to its APIDevelopers using the Responses API can now keep a cached prompt prefix for at least 12 hours instead of the default 30 minutes. A 12-hour cache write costs 2.5 times the normal input price, while reads stay at the usual cached-input rate.
  • OpenAI OpenAI acquires Orchestra Bio, a drug development coordination platformOpenAI has acquired Orchestra Bio, a platform where people and AI agents coordinate drug development programs. Orchestra's team and technology are joining Rosalind, OpenAI's AI models and tools for life sciences research.
  • OpenAI OpenAI publishes coding evaluations for low-resource languages such as COBOLOpenAI introduced MultiLangBench, where a model builds a SQLite-compatible database engine in 20 languages, and a COBOL migration evaluation. On MultiLangBench GPT-6 Astra averages an 87.4% test pass rate against 26.8% for GPT-5.6 Sol, and GPT-6.1 Sol reaches 86.6% COBOL migration accuracy at low reasoning effort.

Usage resets

Completed and reported resets, announcements and banked resets. Only Completed entries count as a finished reset. Full monthly logs: Codex resets in October 2026.

  • Codex Codex reset on Oct 25, 2026 (completed, Oct 24 US time)Day 20 of the 28-day series: posting at 11:06 pm Pacific on 24 October, before the promised midnight PST, Tibo wrote "It is all propagated", reporting the Codex reset as done.
  • Codex Codex reset announced on Oct 25, 2026 (Oct 24 US time)Day 20 of the 28-day series: replying a few seconds after the "Rest and reset" post, Tibo said the Codex reset lands "by midnight PST". The post does not confirm completion.
  • Codex Codex reset on Oct 23, 2026 (completed)Day 18 of the 28-day series: posting at 9:47 pm Pacific on 22 October, a minute after the Day 18 post, Tibo said it lands "tomorrow before 10am PST together with a small ship". It landed on 23 October (UTC), as reset trackers' account monitors recorded; no separate completion post followed.
  • Codex Codex reset on Oct 22, 2026 (completed)Day 17 of the 28-day series: posting at 11:55 pm Pacific on 21 October, Tibo said the reset lands "tomorrow before 10am PST". It landed on 22 October around 10am Pacific (17:00 UTC), as account-monitoring trackers recorded; no separate completion post followed.
  • Codex Codex reset on Oct 20, 2026 (completed, Oct 19 US time)Day 15 of the 28-day series: after an earlier "we're cooked" post, Tibo wrote they are still cooking "and reset. Enjoy.", reporting the Codex reset as done.

Announcements on X

Our one-line summaries of selected posts, each linked to the original. Not the full account timelines.

Claude Code releases (7)

  • Claude Code 2.1.310Added the Monitor tool and plugin monitors on Bedrock, Vertex, and Foundry, and when telemetry or nonessential traffic is disabled; long foreground sleeps are…
  • Claude Code 2.1.309Fixed sandboxed Bash commands on macOS and Linux running the shell's startup files outside the sandbox before each command · Fixed macOS sandbox denyWrite…
  • Claude Code 2.1.308
  • Claude Code 2.1.307Added /claude-api api-version to move Claude API code to API version 2026-10-21 · Added the claude_code.safety_event OpenTelemetry event for each tool call a permission…
  • Claude Code 2.1.306Fixed claude -p inside cloud sessions reporting "Not logged in" when started without the session's environment variables · Fixed an API error being replaced…
  • Claude Code 2.1.305Added sandbox.allowNestedUserNamespaces so sandboxed commands on Linux can create user namespaces (Chromium, rootless podman); this lets them get around the sandbox's file protections ·…
  • Claude Code 2.1.304Added a transcript line when a plugin that runs first skips one of your mod's hooks, naming the event and the plugin · Added…

Claude Code changelog

Codex CLI releases (1)

  • Codex CLI 0.164.0Save reusable prompts and slash commands in ten numbered /presets slots, with a keyboard menu for quick access. (#54380, #54381) · Set model, reasoning…

Codex CLI changelog

Latest news · Usage reset tracker · All releases · Blog

Also tracked: All news · Usage reset tracker · All releases · Blog