Tap Notes: The Extension Was the Attack

Three of today’s items are the same lesson from different angles: the thing you hand your agent (or your agent runs inside) is now the attack surface. Extensions, platforms holding API keys, and the tool definitions agents read are all trust boundaries. Most of us drew them in pencil.
The incident
Backdoor in Claude Code for VS Code extension on the Cursor Marketplace
Backdoored builds of the Claude Code extension (2.1.312 and 2.1.313) were published to the Cursor Marketplace, dropping persistent SSH backdoors across all three major operating systems. DoublePulsar’s writeup covers the same incident from the “how bad is this” angle.
Why it matters: An agentic coding tool is already something you’ve given your shell, your keys, and your repos. Poison the installer and the attacker inherits all of it without needing a single prompt injection. If you run coding agents, find out where each install actually came from (which marketplace, which publisher, which version) before the next incident makes you find out the hard way. “Official” in the name is a label, not a guarantee.
An agentic coding tool is already something you’ve given your shell, your keys, and your repos. Poison the installer and the attacker inherits all of it.Post to X
Northflank’s status page reports a breach, and the discussion points to customer AI API keys being exfiltrated, tied to a wider campaign against cloud platforms.
Why it matters: Platforms that hold your model keys are now worth attacking for the keys alone. Scope them tightly, give each deployment its own, and keep rotation boring and fast. The worst time to learn your key rotation is manual is the afternoon of the breach.
The surface is moving
WordPress AI Team Outlines Four-Part Plan for Native AI Agent Support in 7.3
The WordPress AI Team wants core to be natively agent-compatible: connect your preferred agent without installing the MCP Adapter plugin or hand-rolling authentication. OAuth and the Abilities API are part of the plan.
Why it matters: When the CMS under a large share of the web makes “agent-addressable” a default, every plugin’s capabilities become part of an agent-facing API whether the authors planned it or not. The authentication and capability model is the part to read closely. Plugin authors who describe their abilities well will be easier for agents to use correctly. Everyone else gets used incorrectly.
The CRM I built this weekend doesn’t have a UI (May 2026)
Chris Lema built a CRM as an MCP server with roughly a hundred tools and no UI, operated by conversation. His argument: tool descriptions are the real product and the real documentation.
Why it matters: The detail that carries it is the failure mode. A patch tool that silently drops keys you didn’t include is a schema footgun, and Claude caught it only by reasoning from the schema. That’s luck, not a guarantee. Design tools that are hard to misuse: partial updates by default, refuse ambiguous writes, return the full record after every mutation so drift is visible. His “check for similar existing tags before creating one” instruction is prompt-as-lint, and it’s worth stealing. A bad screen is annoying. A bad tool description corrupts data quietly for weeks.
Thinking about the work
Fragments: April 2 (April 2026)
Martin Fowler’s roundup stacks several pieces into one argument: Margaret-Anne Storey’s three-layer debt model (technical, cognitive, intent), a Wharton paper on AI as “System 3” and “cognitive surrender,” and Ajey Gore’s case that verification becomes the expensive human job.
Why it matters: Intent debt is the new idea. It’s the debt an agent inherits: when the only record of what you meant to build is thin or stale docs, tests, and commit messages, no model is good enough to steer. The surrender-versus-offloading split is also useful because you can catch yourself doing it. Offloading is delegating during deliberation. Surrender is rubber-stamping output because it looks plausible, on work you could have checked.
Drew Breunig baked a working NES 6502 CPU emulator into the weights of Qwen3.5-4B. It plays Super Mario Bros, very slowly. The Opus 5.5 prompts he used are public, and the model found a 15-year-old CPU conformance suite to check against.
Why it matters: Absurd project, serious method. The agent succeeded because a fixed external oracle already existed. That’s Fowler’s intent debt from the other direction: when intent is executable, agents can actually pay it down. When you give an agent a hard task, look for the test suite first.
One more thing
The three extension links above are the same event in different registers. Read the Phorion one for mechanics, DoublePulsar for scale, and the comment threads for everyone discovering which of their machines had it installed.
The basement has opinions about marketplaces now. 🪨