Platforms, Endpoint MCP, and Cory Doctorow
19 Nov 2026
A week after burying myself in stacks of 100-page MCP security guidance documents because I’ve been asked to write an organization some standards, I just listened to the “Marketing Ships to Production” on Arrested DevOps, and it’s re-swirling my recurring thoughts that “platform engineering,” handling today’s likelihood of nondeterministic compute running on enterprise employee endpoints (laptops, etc.), and Cory Doctorow’s 2011 talk “The Coming War on General Computation” are deeply intertwined.
Endpoint agentic GenAI can be scary
If you deeply read enough material about the Model Context Protocol (MCP) from the Center for Internet Security (CIS), the Open Worldwide Application Security Project (OWASP), the Non-Human Identity Management Group (NHIMG), the Agentic AI Foundation (AAIF), the National Security Administration (NSA), and the National Institute of Standards and Technology (NIST), you’ll probably pause for a moment to consider running away screaming.
I don’t think you’d be unreasonable to conclude that “AI coding assistant” desktop applications and application plugins that do nondeterministic compute (like Anthropic’s Claude Code, OpenAI’s Codex, Google’s Antigravity, or Microsoft’s GitHub Copilot) have no business running as/in processes within employee laptops’ operating systems. Heck, not even within those laptops’ “user” process space.
I mean, have you seen what the average employee laptop can do, even without OS admin rights? The files available in its filesystem? The networked devices it can find and potentially authenticate into as long as it’s connected to the VPN? The internet it can reach? The Outlook mailboxes pre-SSO-logged-into in its Edge web browser?
And that’s before you consider making it even easier for the AI coding assistants to muck around by wiring them up to MCP servers, or by letting them do “computer use.”
- Which, by the way, surprise, are both potentially already enabled or enableable-per-employee by default, unless you find the right settings to disable them enterprise-wide. Fun, right?!
- Good grief, the other day, I tried to test what “computer use” could do by asking GitHub Copilot’s desktop app to write “Hello Word” in a new Word document. When it couldn’t, for whatever reason, find the Word application (a peer suggested maybe it forgot Word is called
winwordas an executable), GitHub Copilot goal-seekingly refused to simply quit and tell me it’d failed. Instead, it persisted onward and opened my web browser. There, it hit the jackpot and found me already logged into a Microsoft Office 365 tenant. It took a screenshot to find the button for New Document and click it, then took a screenshot to figure out the right pixel at which to click and type “Hello World.”- Yeeeeeahh. Endpoints are powerful, and we humans have historically left so much general computation security to compute behaving deterministically!
Believe me, I get tempted to shout from the rooftops that the only way to safely enable the velocity that Matty Stratton, Tanya Higgins, and Stephan Schmidt talked about on that recent Arrested DevOps episode is to make ephemeral tightly-restricted sandboxed compute part of your “platform.” To insist that long-lived endpoints like employee laptops or persistent VDIs simply aren’t allowed to be used for developer activities anymore, if “developer activities” is going to rely upon nondeterministic compute to assist their velocity.
But that’s expensive! (Which is why I’m reading the rest of the hundreds of pages, about what else to do before budgets are obtained for all of that.)
Endpoint agentic GenAI reminds me of 2011
Also, though? There’s this little part of me that just … I grieve it.
I grieve the fall of determinism as the default for endpoint compute because it … it lends weight to the kinds of arguments that people fearmonger with. People who, in 2011, Cory Doctorow pointed out often have a significant financial interest in selling restricted-power computation machines, and/or in preventing ordinary humans from using general computation to fight oppression.
- Before, Cory warned us of fearmongering that general computers can do bad things like DDOS water treatment facilities, so only corporations should own machines capable of general computation.
- Now, I find myself fearmongering myself that general computers with AI coding agents installed onto them can do bad things like accidentally delete all of a team’s instruction manuals out of Sharepoint.
And I feel the same old tension of guilt that I might not hit the sweet spot balancing safety and freedom.
It’s like Matty said in that episode this week – none of 2026’s worries are new or unsolved. Generative AI’s just raising them at an overwhelming scale.
Endpoint agentic GenAI bookmarks
I haven’t read most of these, but my chats with various LLMs have, to a pretty substantial degree. And a lot of them are in my bookmarks. I’ve indicated the ones I’ve actually read personally.
Now you, too, can bury yourself in docs, if you’d like.
- OWASP MCP top 10 security issues (brings you to a GitHub repo that’s a little clunky to click through – honestly, I left that to my LLM and personally just listened to the fake podcast in the next link)
- If you don’t have time to read the actual OWASP MCP top 10 security issues, here’s the OWASP top 10 summarized to you that you can pull up in a podcatcher – and yes, it’s one of those AI slop fake podcast episodes, but it did a good job and it’s just 20 minutes
- OWASP MCP security cheat sheet – I liked this one
- CIS controls MCP companion guide (PDF) – I liked this one
- NSA MCP Security Design Considerations for AI-Driven Automation (PDF) – I liked this one
- OWASP LLM06:2025 Excessive Agency
- OWASP Secure Coding with AI Cheat Sheet
- AAIF MCP security best practices
- NHIMG AI Coding Agents Security Guide: Protecting Secrets, Code and Pipelines
- NHIMG MCP Security Guide: Securing the Model Context Protocol
- NHIMG Agentic AI Security Policy Template
- NHIMG AI Agent Observability, Audit and Incident Response Guide
- NHIMG Threat Modelling AI Agents: A Practical Workshop Guide
- NHIMG AI Agent Authorisation Guide: Least Privilege for Non-Deterministic Actors
- NHIMG AI Agent Deployment Checklist: Pre-Production Security Checks
- NHIMG AI Agent Tool and Plugin Security Guide: Tool Poisoning
- NIST SP 800-207: Zero Trust Architecture
- NIST AI Risk Management Framework
- NIST SP 800-218: Secure Software Development Framework