GitLab cut staff and called it "Act 2." Bambu Lab is suing the maintainer of a fork that lets you use your own printer without their cloud. Anthropic's Mythos model found a real bug in curl. Maryland sent a $2B grid bill to ratepayers for out-of-state AI data centres. The Shai-Hulud worm shipped signed malicious TanStack and Mistral npm packages.
Issue 20
15 May 2026
Hi Stranger,
Two of this week’s bigger stories were about who actually owns the thing. Bambu Lab sent a legal threat to someone who built a way to use your Bambu printer without their cloud. GitLab cut staff and called it "Act 2."
A quieter one underneath: Anthropic’s Mythos model found a real bug in curl this week. The maintainer wrote it up like any other useful contribution, not a marketing moment.
Industry
1. GitLab announces a restructuring and calls it "Act 2"
GitLab published Act 2 on Monday – a public letter to customers and investors that opens a voluntary separation window and restructures around "agentic" software engineering. The phrasing is unusually direct: "we grew into a shape that was right for the last era and isn’t right for this one." Q1 and FY27 guidance is reaffirmed. The financial scope of the cuts: held back until the 2 June earnings call.
The strategy half is what’s worth reading. GitLab is betting that the orchestrator role – CI/CD, security, source, ticketing, agent control plane – is what matters once most code is written by agents and not typed by humans. Five rebuilds. No immediate product disruption for anyone running self-managed.
2. Bambu Lab threatens the maintainer of an OrcaSlicer fork
Jeff Geerling wrote up the Bambu Lab story. The company sent legal threats to the developer of OrcaSlicer-bambulab – a fork that lets you use your Bambu printer without routing every print through Bambu’s cloud.
OrcaSlicer is AGPLv3. It descends from Bambu Studio, which descends from Prusa Slicer, which descends from slic3r. The AGPL chain is the thing that’s meant to keep this kind of move out of bounds.
Bambu calls the fork an "impersonation attack." That’s a creative way to describe "you can connect to the printer you bought without our servers in the middle." Louis Rossmann offered to pay the maintainer’s legal fees.
I have an X1 Carbon at home. The printer is good. The cloud default is not. "You bought it, you don’t own it" is starting to annoy me more than I’d like to admit.
3. Internet Archive opens a Swiss branch
The Internet Archive announced Internet Archive Switzerland, a separate non-profit hosted in Zürich that will hold a copy of the archive under Swiss jurisdiction. The framing is preservation. The subtext is the US legal environment. Switzerland is not the EU, not Schengen, and has its own well-settled habits around copyright and digital archives. A second physical site for a single-jurisdiction memory institution is the right kind of hedge.
Artificial Intelligence
4. Mythos finds a curl vulnerability
Daniel Stenberg wrote up the first real find from Mythos – the Anthropic model Anthropic flagged in April as so good at finding security bugs they were releasing it only to selected partners. One vulnerability. Reported and patched. Stenberg writes it up the way he’d write up any other contributor’s report: no marketing beat, no big takeaway, just a maintainer logging a useful contribution.
The number that matters here is per-finding signal-to-noise, not headline volume. If Mythos and its successors keep landing real bugs with low false-positive rates, the running "AI dumps garbage on maintainers" line might not survive 2026.
5. Local AI needs to be the norm
Cyrus on unix.foo argues for local-first AI and gives the example that makes it concrete: his news aggregator summarises articles on-device through Apple’s local model APIs. No server detour. No retention question. No vendor billing surface.
The piece isn’t anti-cloud. It’s anti-default. The default of slapping an OpenAI or Anthropic call onto every feature turns a UX nicety into a distributed system, with consent, retention, rate-limit and uptime risk attached. Most of what developers reach for the cloud for runs fine on the silicon people are already carrying.
Infrastructure
6. Maryland sends a $2B bill to ratepayers for out-of-state AI
Maryland filed a complaint with federal energy regulators over a roughly $2 billion grid-upgrade bill being passed through to Maryland ratepayers to serve AI data-centre load – most of it sited outside the state. The state’s argument is the simple one: ratepayer-protection rules say in-state ratepayers shouldn’t be subsidising out-of-state load growth.
First state-level complaint that names AI build-out specifically as the cost driver. It won’t be the last.
7. Google Cloud Fraud Defence is WEI in new packaging
Private Captcha points out what Google Cloud Fraud Defence actually is: a server-side rebuild of the Web Environment Integrity proposal that died publicly in 2023 after the developer revolt, repackaged as a paid B2B abuse-prevention product. Same technical core. Attest the browser and device. The server decides who gets through. Anyone outside the attested set gets gradually downgraded.
The route around the public spec process is the interesting part. WEI couldn’t ship as a browser standard because of how visibly it would have broken the open web. Shipping it as a paid Google Cloud feature gets to the same place, more quietly, with money on the line that makes adoption sticky.
Microsoft
8. May Patch Tuesday: 120 flaws, no zero-days
The May 2026 cumulative fixes 120 vulnerabilities and ships with no known-exploited zero-days. "No zero-days" is the unusual bit – most months this year have shipped with at least one. A regular maintenance window is fine.
9. Teams will block screen capture during meetings
Microsoft announced a Teams change that blocks screen capture by participants during a meeting. The use case is confidentiality of sensitive material. The trade-off: nobody on the call can keep a recording of what they themselves saw, including for notes and accessibility.
Worth telling people on the call before you flip it. A tenant policy that quietly removes a workflow tends to surface as a complaint two weeks later.
Development
10. The Emacsification of software
Thomas Ptacek argues that we all need better Markdown viewers than we currently have, because the agentic-TUI era has buried everyone under unreadable terminal output. The framing he reaches for is "Emacsification" – the slow drift where editing layers absorb everything around them and turn small utilities into mode-stacked behemoths. Funny about Charm’s glow, kind to his friend Josh’s Markless, serious about the point: at least 14% of LLM-coding burnout right now is people scrolling escape codes that should have been rendered.
11. Going back to writing code by hand
The k10s devlog walked through a multi-week experiment in keeping the developer fully out of the loop while building real software with a coding agent. Their conclusion, verbatim: "I still need to be in the loop to make anything meaningful."
Two takeaways worth borrowing. God-object code is to AI coding what the em-dash is to AI writing – a tell, not just a smell. And vibe-coding compounds in a particular way: every cheap shortcut becomes the next decision’s premise. Closer to the experience of running Claude Code as a daily driver than most "AI is over" retrospectives.
Information Security
12. Shai-Hulud ships signed malicious TanStack and Mistral packages
The supply-chain campaign tracked as Shai-Hulud shipped signed malicious versions of TanStack and Mistral npm packages this week. TanStack maintainers posted a postmortem of how the account compromise played out.
The signature step is the part to keep in mind: npm provenance attestations look the same on a legitimate publish and on a compromised-credentials publish. The attestation is about who pushed the artifact, not whether they meant to. Anyone with TanStack or Mistral pinned around 10–12 May should pull the lockfile and check what landed.
13. Hardware attestation as a monopoly lever
GrapheneOS posted a long argument that hardware attestation – Play Integrity, Apple App Attest, equivalents – is on track to become the next platform-monopoly lever. Banks, payment networks, government services, public-transit apps, and major streaming services already use attestation to decide whether to talk to a device. Devices that fail it – rooted phones, alternative OSes, anything that won’t sign the right way – quietly get shut out of services that have nothing structurally to do with device integrity.
The piece is careful to separate attestation as a fraud-prevention tool (narrow, opt-in, fine) from attestation as the universal default for service access (the trajectory). Worth reading. Same argument that put Google Cloud Fraud Defence (item #7) on the watchlist, one layer deeper.
Coming up
Tue 19 May: Google I/O 2026 keynote (10:00 PT, livestreamed). Two days of sessions on Gemini, Android, the agent toolchain. The keynote is where the year’s Gemini cadence and pricing usually land.
2–3 June: Microsoft Build at Fort Mason in San Francisco. First Build outside Seattle in nearly a decade. AI-and-agents-heavy track list. 2,500-attendee cap.
2–5 June: Computex Taipei. Theme: "AI Together." Qualcomm, Marvell, Intel (Lip-Bu Tan) and MediaTek keynotes anchor the silicon side of the year. Same week as Build – the agentic-AI story splits across two continents.
Plenty to react to. Reply if any of it lingered.
All the best,
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.
Hanso Pte Ltd · 1 Phillip Street #08-00, Singapore 048692