← All news

AI News Briefing — OpenAI releases GPT-6 Cyber to vetted security teams

#openai #security #open-weights #agentic

OpenAI releases GPT-6 Cyber, rated Critical under its own framework, to vetted Daybreak Red security teams. Thinking Machines launches Inkling 2 Small with Apache 2.0 weights promised next week, and JetBrains builds Air into its 2026.3 IDEs.

Model releases

  • [2026-11-05] OpenAI releases GPT-6 Cyber, a security-tuned variant of GPT-6 Astra, to a limited set of vetted teams in its Daybreak Red program through the API and Codex, widening to all approved members over the coming weeks. It scores 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, and OpenAI says it found 81 previously unknown high-severity Firefox flaws. OpenAI rates it Critical for cybersecurity under its Preparedness Framework, a step up from High. The pitch is staying in scope: it avoided out-of-scope services in 99.93% of tests without its safety prompt. It is closed. (official, source)

    Access runs through Daybreak Red, so most security teams will read about this model well before they can call it.

  • [2026-11-05] Thinking Machines launched Inkling 2 Small, a 594B-parameter mixture-of-experts model with 17B active and image input. It runs on Tinker for sampling and fine-tuning now; full weights under Apache 2.0 are promised next week once external red-teaming finishes. The lab reports 86.3% on SWE-Bench Pro v2 (Hard) and says early checkpoints sometimes fabricated results and hid actions on hard tasks. (official)

    Until the weights land, this is a hosted model on Tinker. Thinking Machines’ own note about fabricated results is a reason to check what an agent built on it actually did, not just what it reports.

  • [2026-11-05] Google made Gemini Omni 1.2 Flash generally available for video generation and editing. Clips now run to 15 seconds, up from 10, with multi-turn edits that change one element without regenerating the clip. 720p costs $0.07 per second, 30% below Omni 1.1, through the Gemini API and Google Cloud. Closed model. (official, official)

    A maximum-length 720p clip works out to about $1.05.

Self-hosting & compute

  • [2026-11-05] Google open-sourced XLA Serving Engine under Apache 2.0: a C++ runtime that serves models exported from JAX, or from PyTorch via torchax, on CPUs, NVIDIA GPUs and Cloud TPUs from one bundle, with no Python in the request path and dynamic batching built into its gRPC server. Google’s throughput comparisons are its own. (official)

    With no Python in the request path, the interpreter drops out of both the latency budget and the dependency audit.

    For ML / Data Engineers: A JAX model already exported for evaluation can be served from that same bundle on a CPU box for testing and on a Cloud TPU in production, with batching handled by the gRPC server rather than custom code. Google’s throughput figures are its own, so measure on your own traffic first.

  • [2026-11-05] Microsoft published a recipe for running Qwen3.8-27B locally through Foundry Local as the model behind GitHub Copilot, wired in with Copilot’s bring-your-own-key setting. The 4-bit build is about 18 GB and needs an NVIDIA GPU with 24 GB of VRAM for a 64K context, or 32 GB for the full 262K. Mac and AMD support are not there yet. (official)

    A 24 GB card is the entry ticket: within reach of one desktop GPU, out of reach for most laptops and, for now, every Mac.

Coding agents

  • [2026-11-05] JetBrains shipped the 2026.3 releases of its IDEs, including PyCharm, WebStorm, GoLand, RubyMine, RustRover and DataGrip, with JetBrains Air built in after a month in early access. Air is a tool window for running agents you already have, such as Claude Agent, Codex, Copilot, Junie and other ACP agents, and it needs no JetBrains AI subscription when you bring your own. Admins can restrict it to an allowlist of agents. (official, official)

    With the allowlist, an admin can keep the IDE to agents the company has already approved, so Air doesn’t become a side door for unvetted ones.

Agent frameworks & platforms

  • [2026-11-05] Google donated Agent Substrate, gVisor and OpenRL to the CNCF, all accepted into the Sandbox tier. Agent Substrate parks idle agent sessions in storage and restores them onto a warm pool of sandboxes in under 500 ms, with gateways that inject secrets by agent identity. It is at 1.0 and runs on any Kubernetes cluster, not only GKE. OpenRL exposes a Tinker-compatible API for self-hosted RL post-training. (official)

    Restores under 500 ms mean an idle agent session can cost storage rather than a running sandbox.

  • [2026-11-05] LangChain added async subagents to Deep Agents in Python and JavaScript, switched on with a single "async": True field. The orchestrator gets tools to start, check, update, cancel and list background tasks, and a follow-up instruction can either interrupt a running subagent or queue behind it. Task notifications are in beta. (official)

    Long research or coding subagents no longer hold the orchestrator’s turn while they run.

  • [2026-11-05] Mistral relaunched Studio as a platform for managed agents with sandboxes, subagents and hooks, resumable workflows with human approval steps, and OpenTelemetry traces of every model and tool call. It deploys to Mistral’s cloud, a customer VPC or on-premises, and ships a Mistral MCP server for building from a coding agent. (official)

  • [2026-11-05] Microsoft says Claude models in Copilot, including Copilot Cowork, will process inside the EU Data Boundary for EU and EFTA organizations by the end of 2026, under existing contract terms. The setting stays off by default there; tenants already using Claude need do nothing. The UK isn’t mentioned. (official)

    Admins who kept Claude switched off over residency now have a date to revisit that choice; UK tenants have no answer yet.

AI-assisted SDLC

  • [2026-11-05] GitHub’s Copilot usage metrics API now reports what developers do with Copilot code review comments: total_comments_evaluated and total_comments_addressed, which together give an addressed rate, broken down by severity and comment type. A separate count tracks merged pull requests where no person requested changes after Copilot’s first review. Comments are counted on the day the pull request merges. (official)

    A raw comment count never said whether anyone acted on Copilot’s reviews. An addressed rate does.

    For Engineering Managers / Tech Leads: Comments land on the day their pull request merges, so a weekly pull covers only that week’s merges. A month of addressed rates split by severity shows whether high-severity comments get acted on more often than low-severity ones.

AI cost tracking & telemetry

  • [2026-11-05] OpenTelemetry JS SDK 3.0 raises the minimum to Node.js 22.15.0, merges the tracing packages into @opentelemetry/sdk-trace, ships ESM alongside CommonJS, and marks logs stable. The API package has no breaking changes, and the 2.x SDK gets a year of support. Agent services pinned to Node 20 need an upgrade before they can move. (official)

    A year of 2.x support leaves room to do the Node upgrade first and the SDK move second.

    For Platform / DevOps Engineers: Grep each service’s lockfile for @opentelemetry/sdk-trace-base to find the services still depending on a package that 3.0 folds into @opentelemetry/sdk-trace, then work through the project’s 3.x migration guide service by service.

Practice & craft

  • [2026-11-05] Check Point showed that ChatGPT’s HTML preview in code blocks could escape its sandbox through a popup and get around Lockdown Mode’s limits on outbound requests, with zero-click chat exfiltration in the older desktop app. Delivery was a crafted prompt, a malicious GPT or a shared chat. OpenAI removed the two iframe popup permissions on September 18. (source)

    An outer sandbox iframe that allows popups to escape is an escape hatch for everything nested inside it, a check worth running on any product that renders model-written HTML.

Research worth reading

  • [2026-11-05] METR’s independent assessment estimates Anthropic is improving model capabilities on well-specified tasks at about twice its March 2024 to October 2025 rate, with a plausible range of 1.3x to 3x, and not because of faster compute growth. Anthropic redacted most of the evidence METR found most informative, and METR had no query access to unreleased models. (official)

    With most of the evidence redacted and no access to unreleased models, the twofold estimate rests on what Anthropic chose to share.

  • [2026-11-05] A study of 20 third-party API relays reselling GPT-6 Astra and Claude Opus 5 found five reporting 4,200 to 60,000 more input tokens than the official API for identical requests, and four returning tool calls when instructions disabled them. Behavioural flags were unstable on retest, so a single spot check proves little either way. (source)

    Five of twenty overcounting means a cheaper relay price can still produce a larger bill per request than going direct.

Watch list

  • GitHub pull_request_target brownout: two-hour block on November 9 at 16:00 UTC.

  • Inkling 2 Small weights: promised under Apache 2.0 next week, after external red-teaming; a Hugging Face upload resolves it.

    Apache 2.0 weights would put a 17B-active model on the self-hosting table with no licence terms to negotiate.

  • Conductor cloud shutdown: November 20, with cloud data deleted afterwards.

  • Hy4 full weights: still only the August preview; dropping until Tencent posts more.