Issue 41

9 October 2026

Hi there,

Microsoft is trying to make the Windows PC a place where agents work locally as well as in the cloud. Its containment layer is available now, while local-model routing and the broader Copilot experience are still ahead. OpenAI has started shipping interactive GPT-6 answers, and Anthropic’s free open-source scanner brings an old question back to maintainers: who checks all the findings?

Industry

1. Anthropic launches a cyber programme for critical infrastructure

Anthropic’s Cyber Mission, announced 8 October, has two arms: support for the companies defending power, water and other industrial systems, and a free opt-in vulnerability scanner for open-source maintainers. For the infrastructure programme, Anthropic says it will provide models, on-site engineers and threat research to specialist defenders, rather than directly to every utility. It is a commitment and a set of launches, not evidence that those systems are already safer. The scanner’s separate terms, including unreviewed findings, are below.

2. Atlassian puts OpenAI models into Rovo

Atlassian and OpenAI expanded their agreement on 6 October. GPT-6-family models will power more Rovo agents using Atlassian’s Teamwork Graph across Jira, Confluence and other work; the companies are also exploring deeper Jira workflows for assigning and reviewing agent work. The announcement is a partnership and product direction, not proof that those proposed assignment features are live. It matters because project context and permissions, rather than a fresh chat window, are where enterprise agents will be judged.

Artificial Intelligence

3. ChatGPT starts generating interactive answers

OpenAI began rolling out GPT-6 with Intelligent UI on 7 October. Depending on a question, a reply can include a chart, form, map or small interactive tool assembled from native components as it streams. Paid ChatGPT tiers got the rollout first, with Free and Go following; OpenAI says this changes the Chat experience, not the models behind Work or Codex. The more consequential change is that an answer need not look like a page of prose.

4. Claude Haiku 5.5 reaches Bedrock

AWS made Claude Haiku 5.5 available through Bedrock and Claude Platform on AWS. It is the first Haiku model with effort controls, letting developers trade compute and cost against the task instead of choosing one setting for a whole workload. Bedrock offers regional inference profiles and the usual AWS IAM, CloudTrail and billing controls; Anthropic’s speed and cost comparisons remain vendor claims.

5. OpenAI publishes model-generated mathematics with Lean proofs

OpenAI released mathematical results from an internal frontier model on 6 October, with many proofs formalised in Lean and a repository carrying revision and citation procedures. It also shared ten reasoning summaries and estimates of compute spent, rather than only polished papers. The mathematics still needs independent scrutiny; machine-checkable proofs help with correctness of a formal statement, not whether the statement captures the intended theorem.

Microsoft

6. Windows makes local AI part of the agent platform

Microsoft’s 7 October Windows announcement is a bigger shift than another Copilot feature. It says Microsoft Execution Containers are now generally available on Windows 11 to constrain an agent’s file and network access at runtime. MAI Code 1.1 Flash is being brought on-device, and Windows ML is adding llama.cpp support. GitHub’s HydraFusion is planned to route between local and cloud models in an experimental preview later in October; Copilot’s promised access to local files, actions and models is expected to roll out over the coming months on Copilot+ PCs. Microsoft is also taking pre-orders for RTX Spark machines built for larger local models. Some pieces are available now, while the routing and Copilot experience remain promises; they should not be collapsed into a single shipped product.

Development

7. GitHub brings a model to secret detection

GitHub says its new model reads surrounding code to identify likely credentials, including passwords that do not match a known token pattern. Existing AI-detected password alerts for GitHub Secret Protection and Advanced Security customers switch to the model at no additional charge. AI checks in push protection and Copilot security review are opt-in and will consume GitHub AI Credits; the latter features are still in preview or planned preview. A context-aware scanner may catch a credential a pattern misses, but the cost and blocking behaviour depend on which check an administrator enables.

8. Stacked pull requests are generally available

GitHub made stacked pull requests generally available, with improvements to rebasing, signed replacement commits, merge queues and stack navigation. Automatic merge for stacks is rolling out over the coming weeks. GitHub reports more merged code and shorter time-to-merge among preview users; those are its own comparisons, not a controlled independent study. For teams splitting a large change into dependent pieces, the mechanics matter more than the headline metric: approvals can stay in place when unchanged code is rebased, and each pull request can retain a distinct merge commit.

9. Triage can archive pull requests

Repository users with the triage role can now archive and unarchive pull requests. Archiving closes the pull request and makes its conversation read-only; unarchiving restores comments and reactions but does not reopen it. It gives trusted moderators a way to clear spam, duplicates and abandoned work without granting write access. This changes who can perform a queue-management action, not who can change code.

10. Genkit Dart reaches 1.0

Google’s Genkit Dart 1.0, announced on 8 October, is the first stable release of its open-source framework for AI features and agent workflows in Dart. It offers one API across several model providers, typed flows and local testing tools. The framework is available on pub.dev; production-ready is the project’s description, not an independent assessment. Its appeal is practical for Flutter teams already using Dart: they can keep model calls and typed workflow definitions in the same language as the application.

Information Security

11. OpenSSH plans more frequent security releases

OpenSSH 10.6, released on 6 October, says the project has received more security reports found with AI tools or AI assistance. Many do not prove security impact under a realistic threat model, the maintainers note, but some issues have been independently found by another researcher. The project says it will make releases more frequently rather than batch fixes until the next planned version. The release also begins warning on scp -R, a fragile remote-to-remote copy mode whose shell-quoting assumptions can differ between hosts.

12. Anthropic offers open-source maintainers unreviewed vulnerability reports

Anthropic launched OSS Scanner on 8 October. Eligible open-source maintainers can opt into free periodic model-driven scans. Anthropic says it found more than 29,000 candidate vulnerabilities in six months but manually triaged about 6,000; the new service sends fully model-generated findings and proposed patches without human review to projects that choose the faster route. It will continue its human-verified disclosure process for projects without triage capacity. That distinction matters beside OpenSSH’s warning above: a scanner can increase the supply of reports far faster than a maintainer can validate or fix them.

13. OpenAI traces influence campaigns through fake organisations

OpenAI says it banned two influence operations that used its models alongside conventional tactics: one Russian operation built a purported think tank, while an Iranian one used seven supposed journalist identities to pitch articles to online outlets. In OpenAI’s assessment, their potential reach was higher where material reached real publishers than where it stayed in social-media replies. Its report describes observed activity and its own impact scale, not a complete measure of the campaigns’ effects.

Coming up

16–25 October — Open Source AI Week, San Francisco Bay Area. Its calendar starts with a Hugging Face-hosted opening event and includes an on-device AI hackathon on 17–18 October.
20–21 October — PyTorch Conference North America, San Jose. The PyTorch Foundation’s programme covers training, inference, agents and the wider open-model stack.

Cheers,
Julian

How this is made

Throughout the week I stumble across a mildly unreasonable number of interesting things, and I forward them instantly to the friend or colleague I think might care – sometimes to their delight, sometimes to their annoyance, and often with no context at all. Heartbeat is the attempt to do that a little better.

Every Friday a small agent I built, Honoka, looks through the places where those links tend to leak out: my private email, work email, Matrix, Mastodon, WhatsApp, Apple Messages, Signal, and the faint imprints on the platen of my Olympia typewriter (still not an API, tragically). It sorts, filters, groups and summarises the week, then hands me a draft.

Honoka is guided by a private corpus of things I have written over the last fifteen years, so it can get closer to how I sound in more-or-less official emails and public notes. I still take a pass by hand: remove things, change sentences, check links, argue with the judgement. Whether that is enough is, frankly, the experiment. Every issue has one item written entirely by hand. If you can reliably spot it, hit reply and judge.