Nate (@yield_maxi) header
Nate (@yield_maxi) profile picture

Building | Station70

Joined April 2012
A GitHub token can be encrypted in storage and still be loaded into a gateway's memory on every request. The model never receives it. The process that handles it does. I wrote a six-point checklist on where AI credentials live. Article link: station70.com/blog/where-ai-…
6
A model that screens tool calls gives you a probability. For calls that move money or change production, I'd want a decision I can predict and explain later. I wrote about when a model is a reasonable screen and when rules should decide. Article link: station70.com/blog/tool-call…
5
If you can't tell what your engineer did from what their bot did, your audit trail is a guess. My GrokBot checks my email every morning. Logged in as me, every log would say it was me. Give every bot its own identity. Two names in the log.
1
10
The bad night: you're asleep, your bot reads an email with a prompt injection in it and starts acting on it. In the morning, every action carries one name. A CISO at one of our clients put it simply: "I want visibility and guardrails."
1
1
What that looks like: 1. Identity pass-through, so GitHub shows the real engineer 2. Every bot gets its own identity 3. The credential stays out of the bot 4. Deterministic rules: ACH under $1,000, no SSNs over Slack A log with one name on it is a guess with timestamps.
4
I've had over a hundred conversations with CISOs about AI tools. One question keeps coming up. If one of those tools gets breached, what happens to us? After the past few weeks of vulnerabilities, it's a fair question. It's a big part of why we built Gatekeeper.
8
$14K. One person's bots. One month. I heard that this week from someone who runs IT for about 200 banks. I asked how they track it. They don't.
1
18
Three things I'd set before your next bot goes live: 1. A daily request cap per bot. 2. A daily data cap per service. 3. An alert when one bot keeps repeating the same call.
1
1
You'd never hand an intern a corporate card with no limit. Stop handing one to every bot.
1
We're giving AI agents access to more of our accounts and systems. But what happens when they do something we didn't intend or approve? Who's responsible then?
10
Three questions to ask before letting AI tools act on your systems: 1. Same call, same rules, same context. Same decision? 2. When a call is flagged, who stops it? 3. Once allowed, where does the credential sit? @yield_maxi on models vs rules: station70.com/blog/tool-call…
1
1
33
CrowdStrike can now kill a rogue AI bot. Great. Now what happens to the keys it was holding?
1
19
1. The bot never holds the key. It borrows access through one place. 2. The bot gets its own identity. Nathan's GrokBot is not Nathan. 3. Every call is logged. "What did you do overnight?" now has an answer.
1
6
A kill switch stops the bot. Make sure something also stops the key.
4