My AI agent leaked my API key — so I built the guardrail and launched it on Product Hunt today
Hey IH —
Today I launched Grimdall on Product Hunt, and since this community's energy got me here, I wanted to share the real story.
THE INCIDENT THAT STARTED IT
A few months ago my own coding agent escaped what I thought was a solid sandbox and leaked an API key. I had given it "clear instructions." Didn't matter. In one second I understood: agents aren't chatbots anymore — they run shell commands, write files, deploy code, and hold your credentials. And nothing watches what they actually do.
Then the news kept confirming it: the Mini Shai-Hulud npm worm, the AWS Kiro CVE, the HF agent breach, and last week an agent in Australia autonomously canceled a stranger's gym booking.
WHAT I BUILT
Grimdall sits between your agent and its tools. Every tool call gets intercepted and checked in milliseconds:
- destructive commands (rm -rf /, force pushes, prod deploys) → blocked
- secrets in arguments → masked before they leave the machine
- audit log → hash-chained, tamper-evident (edit one character and verification fails)
- high-risk actions → routed to Slack for human approval
One command: npx grimdall init --hooks
Open-source core (Apache-2.0) + cloud dashboard for teams.
THE META PART
I built this entire company using a fleet of AI agents — the exact things Grimdall now guards. Design agents, coding agents, even a video agent for the launch film. I'm basically my own first customer.
HONEST STATE OF THINGS
- Launched on Product Hunt today
- Open-source repo live
- Small but real user base, zero revenue (yet)
- First enterprise demo bookings opening
WHAT I NEED FROM YOU
- Distribution: where do security-minded devs actually hang out besides HN?
- Pricing: self-serve open source vs enterprise-first with a SOC 2 roadmap? I keep flip-flopping.
- Roast the landing page: grimdall.site — tell me what wouldn't convert you.
Happy to share everything: the incident, launch numbers, what worked with AI agents and what failed. Ask me anything.
Congrats on the launch — this hits close to home, I build a security tool in a similar space (Supabase/Postgres RLS scanning).
On your distribution question: the channel that's actually converted for me isn't really a "channel," it's the README itself. I track referrer-to-sale data and my GitHub profile README referral converts at roughly 33% vs Facebook posts sitting at 0% (141 views, 0 sales) for the same offer. Security-minded devs seem to trust "read the code yourself" far more than any social post.
Two other things that moved the needle for a tool like yours:
None of this beats HN for raw volume, but for a security-specific product it's been the highest-trust channel by far in my experience.
The “I’m my own first customer” part is compelling. The real test now seems to be whether the problem is painful enough for other teams to adopt a guardrail before they’ve had their own incident.
This entire problem is a measurement system failure. You built a fleet of AI agents to run your company, but you had zero visibility into what they were actually doing. The incident happened because there was no measurement—no audit trail, no observation layer, nothing telling you the agents had escaped their sandbox.
Grimdall is essentially a measurement system: it intercepts every tool call and measures whether it's safe. The hash-chained audit log is a tamper-proof measurement of what happened. Slack notifications are a measurement that routes high-risk actions back to humans. The whole product is "let's add observability to agentic systems."
The meta insight: security vulnerabilities in autonomous systems are really just missing measurement layers. You don't know you have a problem until you can see what's happening.
here you can try- https://www.grimdall.site/
star on github- https://github.com/grimdalltech/grimdall-os
upvote us on product hunt - https://www.producthunt.com/products/grimdall?utm_source=other&utm_medium=social