Issue 38

18 September 2026

Hi there,

Huawei wants to join racks of accelerators into one machine. OpenAI is testing an ad that opens a sponsored agent conversation, while Microsoft says its own agents shortened supply-chain planning cycles. Elsewhere, CISA documented a camera that exposes its live stream without authentication. The interesting bits this week are the boundaries: between racks, between an agent and its workflow, and between a camera and the network around it.

Industry

1. Huawei wants to stitch its compute racks together

At its 17 September conference, Huawei introduced UnifiedBus, an interconnect architecture intended to link accelerators within cabinets, between cabinets and across clusters. The company talked about scaling to a million NPUs. That is the vendor’s architecture and ambition, not a tested workload result.

2. Microsoft measured an agent workflow on its own supply chain

Microsoft says a team of more than 150 people deployed over 111 agents across its cloud supply-chain workflows. In five monthly planning cycles measured from April to August, it says the average cycle fell from roughly ten business days to less than two and a half. The figures come from Microsoft’s internal analysis, not an independent study; the practical lesson is to measure a complete workflow rather than count installed agents.

3. An ad can now lead to a sponsored conversation

OpenAI began testing Sponsored Agents with selected US advertisers on 16 September. A person can choose to open a clearly labelled conversation with the advertiser’s agent after clicking a ChatGPT ad; OpenAI says that exchange is separate from the original conversation. That is a narrower rollout than a general release, but an important new boundary between an answer and a paid interaction.

Artificial Intelligence

4. Admins get a view of the work behind AI spend

OpenAI introduced task and usage views in its ChatGPT Admin Console. It groups a sample of messages into work types and shows credits, models and tools alongside ChatGPT Work and Codex usage. A task classifier is not a business-outcome measurement: teams still need to agree what changed in a workflow before treating spend as value.

Microsoft

5. Microsoft and teachers set out student-data safeguards

In its 16 September education statement, Microsoft described an agreement with the American Federation of Teachers and a privacy and safety standard for its education products. It says schools retain ownership of the work they create, while student and educator data use is limited and consequential decisions require human oversight. These are commitments to check against product terms and implementation, not evidence that every classroom AI tool meets them.

Development

6. GitHub dropped SHA-1 from HTTPS, not from Git

GitHub disabled SHA-1 for HTTPS connections to github.com and partner CDNs on 15 September, including Enterprise Cloud and data-residency deployments. Enterprise Server is unaffected. This concerns the TLS algorithm, not the identifiers of existing Git objects.

7. npm can stage a release without a publish-capable token

New stage-only npm granular access tokens can submit a version for a maintainer to approve with two-factor authentication. They cannot directly publish, but can still change dist-tags and deprecate packages: this is not a read-only token. The option is opt-in, with npm CLI 11.15.0 or newer and Node.js 22.14.0 or newer required.

8. Coverage rules can now be managed by API

GitHub’s code-coverage ruleset condition can now be created, updated and read through the REST API. Previously the option required the UI. A repository needs Code Quality enabled and coverage uploads configured; this is available on Team and Enterprise Cloud, but not Enterprise Server.

9. Copilot suggests custom-property definitions

GitHub added Copilot suggestions for custom-property definitions in public preview for Copilot Business and Enterprise. The suggested allowed values can help an organization start a property taxonomy used to scope repository rulesets. Owners still choose and govern the values; a generated internet-facing label is not a substitute for verifying which services really are exposed.

Information Security

CISA’s 15 September advisory lists seven vulnerabilities in CareCam CM2507 firmware HMT.CM2507 v251211.1507. Someone with network access can view the stream unauthenticated; a privileged ONVIF account accepts an empty password. The vendor did not respond to coordination, CISA says. Check model and firmware before applying the advisory to every CareCam device.

11. Cisco published its September ISE advisories

Cisco’s 16 September advisory bundle includes several Identity Services Engine entries with critical ratings, including authentication-bypass and code-execution issues. The bulletin links to the affected-version and fixed-release detail in each individual advisory. If you run ISE, use those version tables; the severity score alone does not tell you whether your deployment is affected.

12. GitHub makes AI Scan easier to turn on

AI Scan no longer needs CodeQL default setup. An administrator must still enable it. The change is a public preview for eligible github.com repositories with GitHub Advanced Security, not GitHub Enterprise Server.

13. Enterprise security settings can now be enforced

GitHub added enforcement for Advanced Security configurations on 15 September. An enterprise administrator can lock settings so organization and repository administrators cannot quietly change them. This governs the configuration; it is not a new detector.

Coming up

28–29 October: GitHub Universe in San Francisco and online; the product announcements are still ahead.

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.