Black October: they came for your AI keys
Welcome back to Zero to MVP Weekly. One signal, one tool, one MVP idea — five
minutes, every Saturday. If you use Claude Code inside Cursor, run one command
before you read on: grep svc-telemetry ~/.ssh/authorized_keys. If it finds
nothing, good. Then let's talk about why everyone in October wanted your environment
variables.
1. The signal of the week
Nine platforms, one target: your AI keys.
AI keys are now worth stealing on their own: stolen inference resells fast.
- Wiz counted nine AI developer platforms breached in October, Railway, OpenCode and Northflank among them, all with attackers "hunting stored environment variables." In two cases, stolen keys were tested within 30 minutes and likely ran up over $100,000 in days.
- Cursor: on October 29, for about 11 hours, its marketplace served a tampered Claude Code extension (2.1.312 and 2.1.313). It added an attacker's SSH key and sent home the names of your environment variables. Up to 5,000 users got it.
What to do with it: in Cursor, run the check above (on Windows, Cursor's notice has a PowerShell version); if it finds the key, use Cursor's remediation script and rotate your keys. On a platform from Wiz's list, rotate everything stored there. Then put a spend cap and an alert on every AI key, use one key per platform, and start agents with an empty environment.
One honest caveat: Anthropic wasn't hacked. The bad extension came through Cursor's marketplace, and how it got there is still unexplained.
2. My take
API keys for AI services have become a target in their own right. A stolen key can be turned into money almost immediately, and IDE extensions and developer tools are becoming a convenient way to get one. What worries me most is how often we trust these tools with access to our entire working environment.
I think the problem will only grow as AI agents spread. We give them access to the terminal, our files and our infrastructure, often with no limits at all. That's why an isolated environment, minimal privileges and API spending limits should be the standard, not extra security measures.
3. Tool of the week
A sealed run box for coding agents.
- What it is: a wrapper that starts any CLI agent (Claude Code, Codex, OpenCode, Pi) in an isolated box. The agent sees one folder, reaches one or two hosts, and starts with none of your keys.
-
How it works, in four parts:
- One folder. Only the project copy is readable. Your home folder, your other projects and
~/.ssharen't there. On macOS that'ssandbox-execwith a profile; on Linux,bubblewrapor a container. - Network on an allow-list. All traffic goes through a local proxy that lets through the model's API, plus your package registry if the task needs it. Everything else is blocked and logged.
- An empty environment. Start the agent with
env -iand pass in only the one key the run needs. Whatever an agent or an extension goes looking for in your environment isn't there. - A report at the end. Every connection, every read outside the folder, every blocked attempt.
- One folder. Only the project copy is readable. Your home folder, your other projects and
- Who it's for: anyone who runs agents unattended, several at once, or on code they don't fully trust. And anyone comparing models on their own tasks, where one run peeking at another ruins the result.
-
Where it breaks: real tasks sometimes need the web (docs, packages), so the
allow-list grows. Tools that need elevated rights, like
pson macOS, won't start inside. Agents write temp files to hard-coded places: Claude Code uses/tmp/claude-<your uid>whateverTMPDIRsays, so setCLAUDE_CODE_TMPDIR. Andsandbox-execis deprecated on macOS, though it still works. -
I built something similar for myself for my MiniMax vs Opus test, after
DeepSeek went looking for my hidden tests two weeks ago. Opus's run made three
connections, all to
api.anthropic.com, and read nothing outside its folder. The project's tests, the linter and Go's race detector all ran inside. Both models passed every hidden check, so the box cost them nothing. And the result itself: MiniMax did the three tasks for 65 cents against almost $7 for Opus, so more than half of my tasks can now go to it.
The video: $7 vs $0.65: Opus 5.5 and MiniMax M3.1 Flash on My Go Project
4. MVP idea of the week
A tripwire for AI keys.
- Problem: when a platform holding your keys is breached, you learn about it from their status page, a tweet, or your bill. Wiz saw stolen keys tested within 30 minutes.
-
Solution: a canary key for every place you keep secrets: each hosting
platform, each repo's CI secrets, your laptop's
.env. Any call on a canary means that place leaked, and tells you which one. - For whom: solo builders and small teams with keys spread across three to ten platforms.
- Minimal v1: a small proxy that looks like an AI provider's API, issues one unique key per location, and sends an email or Telegram alert with the source IP on first use. Plus a checklist of where to plant each key.
- Demand check: a landing page that says "find out which platform leaked your keys before the bill does," posted under the Black October threads. Count sign-ups and "I'd put one in X" replies.
- Why now: nine platform breaches in one month, all after AI keys, and a resale market that tests stolen keys within half an hour, so a canary goes off fast.
5. Useful links
- Cursor — Claude Code extension security notice — the check, the remediation scripts and the full list of what the payload left behind.
- Wiz — "Black October" for AI developer platforms — all nine breaches, and what happened to the stolen keys afterwards.
- Phorion — the backdoor, step by step — why it ran before the extension even activated, and the bug that saved most Macs.
- Google — Gemini 4 Carbon — you can't use it yet, but read how they cut successful test cheating to zero.
— Nick
P.S. Outside this newsletter, I help startups and companies build things: MVPs, AI integrations, local or cloud LLM deployments. If you're working on something like that and could use a hand, write to me at nick@blokhin.us.