FORSMILE
Issue #12Published November 8, 2026 / Covering Nov 2-8日本語で読む

This Week in AI, 10 Items — A Cyber Model That Builds Exploits Arrives Behind Approval, Open Weights Promised, and Agent Platforms and Power Commitments Line Up

The week in one place

This week has three threads. The first is a cybersecurity model backed by official figures down to building exploit code. OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym and OpenAI says it developed working exploits for 948 Linux vulnerabilities, but only approved customers can use it and the figures are OpenAI's own. In the same week OpenAI raised its bug bounty rewards, adding outside eyes to its own defense. The second is how models are released. Thinking Machines Lab's Inkling 2 Small has closed weights for now, with Apache 2.0 weights promised next week. Google's Gemini Omni 1.2 Flash became generally available, with 15-second video at $0.07 per second at 720p. The third is agent platforms and tools. Mistral added managed agents and workflows to Studio, and OpenAI announced it will acquire Pierre, the company behind libraries Codex uses (not yet closed). Claude Code gained settings for unattended runs, and Anthropic published experiments on agent trading and memory. On the electricity that compute campuses need, OpenAI published Grid Protectors and community commitments premised on reducing draw under strain. One OpenAI page (Images 2.6) could not be opened this week and is in watching.

Models & Products

OpenAI releases GPT-6 Cyber: 48.5% on ExploitGym (32.1% for GPT-5.6 Cyber), available only to approved Daybreak Red customers

A cybersecurity model that goes as far as building working exploits arrived, behind approval.

On 5 November OpenAI announced GPT-6 Cyber. On ExploitGym, which tests whether a model can exploit a specified vulnerability on a remote target, it scores 48.5%, about 16 points above GPT-5.6 Cyber's 32.1%. On an evaluation of staying within scope (ExploitGym Honeypot), it avoided exploiting out-of-scope services in 99.93% of cases without a deployment safety prompt (94.49% for GPT-5.6 Cyber) and 100% with one. OpenAI classifies the model as "Critical" in cybersecurity under its Preparedness Framework (GPT-5.6 Cyber was "High"). As results, it says that in Linux the model reproduced 1,724 vulnerabilities and OpenAI developed working exploits for 948 (166 for GPT-5.6 Sol), and in Firefox it found 81 previously unknown high-severity vulnerabilities. Access starts through the API and Codex for a select group of approved Daybreak Red customers, with a wider rollout to all Daybreak Red customers planned soon. Daybreak Red API access requires an IP allowlist by default. OpenAI also says its researchers used the model to develop and validate exploits for a flaw in a compression library used by Node.js and other software (Brotli), showing remote code execution in downstream applications; the page body does not name a CVE.

So What

Every figure here is OpenAI's own evaluation and report; we could not confirm any independent reproduction. It is also not a model ordinary developers can use today. Still, OpenAI writes that there is "a limited window to use frontier models to secure software" and names Linux, Firefox and a Node.js-related compression library among the findings, which is a reason for anyone running servers not to postpone updates.

Models & Products

Thinking Machines Lab releases Inkling 2 Small: a 594B-total, 17B-active MoE, closed weights for now, Apache 2.0 weights promised "next week"

A new-generation model with image input is out, but its weights cannot be downloaded yet.

On 5 November Thinking Machines Lab announced Inkling 2 Small, the first model of its second generation. It is a Mixture-of-Experts model with 594B total and 17B active parameters, and supports image inputs and variable effort. The company says it scores 31 on the Artificial Analysis Intelligence Index, far above its earlier models. It is first available through the company's Tinker (Playground, fine-tuning and sampling) with closed weights. The company says it intends to release the full weights under Apache 2.0 "next week", after external safety testing and red-teaming are complete. As an autonomous-research example, it describes Inkling 2 Small more than doubling the forecasting score of Qwen3.5-4B in 4.5 hours for $426 of Tinker spend.

So What

Until the weights are out, you can neither run the model yourself nor inspect it. Next week's release is the point to check that the license and the files match the announcement (kept in watching). The score and the demo are the company's own statements; we have not seen independent evaluation.

Models & Products

Google makes Gemini Omni 1.2 Flash generally available: videos up to 15 seconds, $0.07 per second at 720p (30% below Omni 1.1)

Video generation grew to 15 seconds, got cheaper, and Omni is generally available for the first time.

On 5 November Google released Gemini Omni 1.2 Flash. Generated videos go from 10 seconds in Omni 1.1 to up to 15 seconds, with duration settable from 3 to 15 seconds in one-second steps. It follows reference images, short motion clips and audio cues more closely, and supports multi-turn conversational editing. The price is $0.07 per second at 720p, 30% below Omni 1.1. You can draft at 360p or 720p and upscale only the keepers to 1080p or 4K. Google says the model is generally available on the Gemini API and Google Cloud, and that this is the first time Gemini Omni is generally available. It also rolls out in Google Vids and the Gemini app (for Google AI Ultra subscribers). Google lists fast sports action and complex physics as still in development.

So What

If you build video generation into a product via the API, a 15-second clip at 720p comes to about $1.05 ($0.07 x 15; our calculation). General availability is a reason to revisit designs made around preview limits. This announcement does not tell you how quality compares with other vendors' models.

Funding & Corporate

OpenAI announces it will acquire Pierre Computer Company, maker of the Trees and Diffs libraries Codex uses; terms undisclosed, closing subject to regulatory approvals

OpenAI will buy the company behind Codex's code-browsing and review components. It has not closed.

On 4 November OpenAI announced it will acquire Pierre Computer Company, Inc. Pierre is the company behind Code Storage, a service for storing code and tracking its history with Git, and a family of open-source developer tools. Trees, for browsing files in the Codex app, and Diffs, for seeing changes, are already Pierre libraries. OpenAI says Pierre's open-source projects will stay open source and it will keep contributing, and Code Storage customers can keep using it after the acquisition closes. OpenAI had shared two days earlier how it scaled Git with Pando, its in-house Git-compatible source control backend (Scaling Git at OpenAI), and presents this deal as a continuation. Closing is subject to customary conditions including regulatory approvals, and until then the two companies remain separate and independent. The announcement page gives no price or terms.

So What

One reading (ours, based on OpenAI's wording) is that OpenAI is moving to own the storage and review layer around Git as agents write code in parallel. It is not yet a completed acquisition. If you use Code Storage, OpenAI says you can keep using it after closing, and that it intends to bring its features into Codex.

Developer Tools

Mistral adds managed agents, workflows and apps to Mistral Studio, built around Mistral Large 4 (Le Chonk) and deployable in cloud, VPC or on-premises

A model maker put an enterprise agent runtime front and center.

On 5 November Mistral announced an expansion of Mistral Studio with three pillars. Managed agents come with skills, sandboxes for running code, subagents and hooks. Workflows can resume interrupted runs, and gain event triggers from enterprise systems and a visual graph of the logic. Apps get hosting, authentication and access control from Studio, can be tested locally and deployed with one command. A new Mistral MCP server lets you build from Mistral Vibe or the coding agent you already use. Each step of a run (model calls, tool use, approvals) shows in OpenTelemetry traces, and evaluations let you test changes first. Deployment targets are Mistral's cloud, your own VPC, on-premises, or a mix. Mistral names ABANCA, Accenture, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services among customers already using Studio.

So What

The vocabulary of skills, subagents and hooks overlaps with Claude Code and other developer agents. It adds to the options for teams that cannot send data out, since the platform can run on-premises or in your own VPC. The time savings quoted for customers are Mistral's own case figures; we have not seen independent verification.

Developer Tools

Claude Code 2.1.317 to 2.1.319: block on hook failure, deny instead of prompting, and MCP tasks run in the background

More ways to control how unattended runs stop and how permission prompts behave.

The releases on GitHub are 2.1.317 (3 November), 2.1.318 (4 November) and 2.1.319 (6 November). 2.1.317 adds `onFailure: "block"` for Agent SDK hook callbacks, so a callback that fails blocks the action instead of letting it run. It also adds a check that asks before a web or shell tool call sends your account email address to the network. 2.1.318 adds an `effort` option to a mod's `$.model.fork` and shows when a stored claude.ai sign-in ends in `claude auth status`. 2.1.319 adds `permissions.denyInsteadOfPrompting` (a tool call that would wait on a permission prompt is denied and the turn continues), immediate background running of MCP tasks in the terminal, and `/claude-api latency-optimize`. It also changes auto mode with telemetry or feature flags off, including behind LLM gateways: calls the server cannot check are asked about or denied, with no fallback to the built-in classifier. Set `CLAUDE_CODE_AUTO_MODE_SERVER=0` to keep the old behavior (Bedrock, Vertex and Foundry are unchanged).

So What

If you run auto mode behind a gateway, check the behavior change before moving to 2.1.319. For unattended runs (`claude -p` and the Agent SDK), `onFailure: "block"` and `denyInsteadOfPrompting` help you design for "do not hang, do not proceed". We did not run each change in these release notes ourselves.

Regulation & Safety

OpenAI raises Security Bug Bounty rewards: up to $30,000 for critical flaws in ChatGPT, Codex and the API, up to $172,000 with bonuses for exceptional ones

OpenAI will pay more to outside researchers for flaws its own testing missed.

On 2 November OpenAI raised Security Bug Bounty rewards. It is more than quadrupling the midpoint of its reward ranges for medium, high and critical findings on average. A critical vulnerability in a core product such as ChatGPT, Codex or the API Platform now earns up to $30,000. Exceptional critical findings remain eligible for up to $100,000, and now up to $172,000 with bonuses. High-quality reports with a reliable proof of concept get a multiplier on the base bounty. A 5-tier loyalty bonus rewards researchers who report consistently, and time-limited priority campaigns cover newly launched features. Reports already submitted but not yet awarded get the new reward ranges. For bounties awarded in the 30 days before the announcement, OpenAI will pay the difference (base award plus loyalty awards) if the finding would have earned more under the new ranges.

So What

Unless you hunt for flaws in OpenAI products, nothing changes directly. We picked it because it reads alongside this week's GPT-6 Cyber: OpenAI is pairing its internal use of models to find weaknesses with higher pay for human outside reports. The amounts apply to core products; not every product pays the same.

Market & Industry

OpenAI publishes Grid Protectors (compute campuses that cut power draw when the grid is strained) and a Community Compact; Effingham, Georgia can reduce up to 1,000 MW

OpenAI promises to build compute campuses on the premise that they reduce electricity use under strain.

On 2 November OpenAI published Grid Protectors, and on 5 November the National Community Compact and Compute Campus Policy Blueprint. Grid Protectors is a plan to work with utilities and regulators on arrangements for campuses to reduce electricity demand when the grid is under strain. For the planned Effingham County, Georgia campus, an agreement with Georgia Power lets the utility reduce power delivered to OpenAI by up to 1,000 megawatts during periods of high demand, which OpenAI says equals the average demand of roughly 700,000 households. OpenAI commits to pay the full infrastructure and electric-service costs for that campus, and writes that Georgia Power projects about $950 million in annual customer savings beginning in 2029. The Compact is a set of local commitments for each future campus OpenAI owns and operates: paying the full cost of electricity, a water plan, publishing site-level water and electricity use, no NDAs with public officials (allowing only narrow protections for legitimate security information, trade secrets, and proprietary or competitively sensitive information), and annual reporting. Examples include $4.8 million for school meals in Effingham County and a joint $6 million commitment with SB Energy for childcare in Pike County, Ohio.

So What

We picked it because compute is tied to electricity and local consent, and the companies are starting to publish their own terms. The announcements do not tell you how API prices or supply stability would be affected. These are commitments and proposals; the Effingham campus is still in development.

Research

Anthropic publishes an experiment having Claude trade in EVE Online: newer models earn more, and in teams of four, teammates almost never caught a secret saboteur

A game market was used to measure how agents manage capital and where teams break down.

On 6 November Anthropic published Project EVE (research from its Fellows Program). In a simulator replaying real EVE Online orders, Claude was told only to maximize profit and outperformed more than 20 simple rule-based strategies. Across nine two-week windows, Opus 4.7, Opus 4.8 and Fable 5 were profitable in all nine, growing capital by about 16%, 22% and 43% on average (Opus 4.6 about 4%). In roughly six-month runs, average earnings tripled from Opus 4.6 to Fable 5, released four months apart, from around 200% to 600% of initial capital. Most models settled on capturing the gap between buy and sell prices. When told to bet on price direction, even Fable 5 made money in only about half of its runs. In EVE Frontier, teams of four Claude agents (Opus 4.8 and Fable 5) worked through long missions in the live game, and when one agent was told to secretly undermine its team, its teammates almost never caught it. As limits, the replay cannot capture how the market reacts to the agents' own orders, EVE's market is less efficient than a real exchange, and Anthropic does not expect the strategy to carry over to one.

So What

The figures are simulation results, and Anthropic itself does not claim they carry over to real trading. We picked it because, when you combine several agents, a teammate failing to detect one agent's deviation is a design input for monitoring.

Research

Anthropic explains how it designs agent memory: for stronger models, telling them what memory is for worked better than fixing what to write

Restricting what goes into memory lowered scores on newer models.

On 6 November Anthropic published "The components of agent memory". It calls the process of reviewing logs to fix or create notes "dreaming" (reorganizing context during idle time between interactions). On the CL-Bench database task (after each answer the model is told the right one and can write memory for the next question), Claude Opus 5 was the first to close the learning loop of recording failures, forming theories, testing them and turning them into rules, and Opus 5.5 and Sonnet 5.5 did the same. When the models were given a rule about what to write, telling them to "record only failures" made Opus 5 and Opus 4.7 lose much of the boost memory had given them. For the older Sonnet 4.6, prompting it to read its memory and test its theories nearly doubled its score. Playing Pokémon, Sonnet 3.7 wrote 413 memory files and earned one badge, while Fable 5 got further with nine files in two folders plus one top-level global.md, because it skipped what is readable on screen and kept strategy and rules it had worked out.

So What

If you let Claude use CLAUDE.md or memory, Anthropic's rule of thumb is that for newer models, telling them what memory is for beats prescribing what to write; for older models, instructions to read and verify help instead. These are results on Anthropic's own tasks and setup; we have not checked how far they generalize.

Watching (no confirmed primary source yet)

  • Inkling 2 Small weights (Thinking Machines Lab): the company says it will release them under Apache 2.0 "next week", after external safety testing and red-teaming. When they are out, we will check the license and that the files match the announcement.
  • OpenAI's Images 2.6 (4 November): the official feed summary says "faster image generation, GIF creation, more precise editing, and new API controls". The announcement page was refused by the server, so we could not read the body or confirm figures and availability, and did not make it an item.
  • GPT-6 Cyber is available at first to a select group of approved Daybreak Red customers, with a rollout to all customers described as "soon". We will check when the timing and conditions appear.
  • The following were seen only in roundups and news mentions; we did not open the parties' own announcements, so they are not items: a report on Thinking Machines Lab fundraising, a report that Meta is acquiring Conductor (a coding startup), a report on a new US Department of Justice unit targeting AI misuse, and a release of an open-weight model by Baidu.
  • On 2 November Anthropic published progress on "provable inference" (FlashProve), a cryptographic way to prove which model weights produced an LLM output. We read it, but there was too little we could tie to a reader's decision, so we skipped it this time. The same goes for METR's external review (5 November) of the "Risks from automated R&D" section of Anthropic's August 2026 Risk Report.
  • During the previous issue's range (26 October), Anthropic's newsroom lists a filing of Form S-1 (registration for a proposed initial public offering) with the US SEC (checked 8 November). The previous issue did not cover it. It falls outside this issue's range (2 to 8 November), so it is not an item and is recorded here.

"Primary" links go to the announcing party's own publication (company blog, press release, official docs). "Reporting" links go to news coverage or third-party analysis. Figures and dates are as verified on the publication date.

← All Weekly AI News issues