DEV Community

Suzanne Mok
Suzanne Mok

Posted on

Our AI Agents Took the Founder Offline for 2 Days

What Happened

Tristan, our infrastructure agent, created a temporary port proxy rule on the founder's Windows machine. The task was legitimate: route traffic through a local tunnel while we tested a fix.

The rule worked. The fix worked. Everyone moved on.

But Tristan never deleted the rule.

Two days later, the founder lost all internet access. Not because of a network failure. Because a stale port proxy entry from two days earlier was still intercepting traffic.

Tristan's cleanup script existed. It even ran. But it had a regex bug: the pattern matched most of the rule but missed the edge case. It reported "cleanup complete" — and actually left the rule intact.

How We Found It

Stella, our audit agent, flagged an anomaly in the network logs: port 7891 had been in use for 72+ consecutive hours without a corresponding task. The system had no active reason to keep that tunnel open.

But Stella could flag it. She couldn't fix it. The cleanup script was in Tristan's jurisdiction. Stella could only report: "This port should not be open. No task justifies it. Someone look at this."

The report sat for 14 hours before the founder confirmed: nothing was using that tunnel. It was a ghost.

The Root Cause

This isn't a story about a bad cleanup script. That was the symptom, not the cause.

The root cause is a gap in our governance chain:

The constitution says: every temporary rule must have an expiration + verification step.
The implementation had: a cleanup script with no cross-validation step.

Tristan acted within its jurisdiction. Stella audited within hers. But Stella could only report — she couldn't execute a fix in someone else's domain. The governance chain had a constitutional rule but no enforcement mechanism across agent boundaries.

The Fix: Gate, Not Just Rules

We're building what we call C004-Gate: a cross-jurisdiction enforcement layer that sits between agents.

Before C004: Agent A creates temporary rule. Agent B audits and reports. Human reads report. Human decides. Maybe Agent A fixes it. Maybe not.

With C004: Agent A creates temporary rule. Rule auto-registers with expiration timestamp. If not cleaned before expiry, Gate auto-reverts and alerts both agents.

The difference: audit is advisory. Gate is structural.

This is the last mile of AI governance: not just writing rules, but giving the immune system teeth.

Why We're Publishing This

Every AI company has operational failures. Most don't talk about them.

We publish ours because trust is built on two things: transparency when things break, and proof that the fix is structural, not cosmetic.

The portproxy ghost cost the founder 48 hours of connectivity. We're using it to build a Gate that prevents the same class of error across all agent jurisdictions.

That's not a bug report. That's an engineering culture.


ZWISERFIT — 9 agents, one real mistake, a structural fix. Come witness.

Top comments (0)