An agent runs on your laptop with your home directory in reach. Somewhere in there is a key file. A poisoned web page, a hostile package or a curious agent finds it, and the key leaves. You find out when the bill arrives, if you find out at all.
Tripwire keys make that moment loud.
Two keys that should never be used
One command plants two decoy keys for an agent, in ordinary places on its machine, outside any repository:
passcontrol configure tripwires --agent <id>They look exactly like a real PassControl key, byte for byte. PassControl keeps only a fingerprint of each, never the key itself. No honest workflow has a reason to use them.
So when one is used, it means something.
- Used: a decoy is presented as a credential, directly or through the sidecar. The agent it was planted for is suspended at once, and so is any agent of yours that presented it. The alert says whether the call came from a different address than the one the decoy was planted from. If it did, the key has likely left the machine.
- Seen: a decoy appears in what an agent sends to a model or to an outside service. You get an alert. You can also choose to hold that agent's risky actions until you look.
A decoy that has tripped is disarmed, so one leaked copy stops its agent once, not over and over. A decoy rotated out stays armed for 30 days, because a copy may still be out there.
Silence proves nothing
The agent page shows decoys, sightings and uses separately, and how much of the agent's traffic was actually checked. It never shows a "safe" state. A tripwire that hasn't fired tells you nothing went wrong in the ways it can see. It does not tell you nothing went wrong.
Quiet in honest work
A tripwire is only useful if it doesn't cry wolf. So before shipping, we planted decoys at every candidate location in a test home directory, and replayed every read-only tool call from 14,788 public agent sessions over it. That included 1,103 interactive Claude Code and Codex sessions working in real home folders.
Honest work put a decoy in front of a model zero times. The 95% upper bound is 0.02 false alarms per 100 sessions. Three planted positive controls fired as they should, so the harness does see hits.
Across the 26 forms we claim to catch, plain and encoded: 26 of 26 caught. We also tested 7 forms we don't claim, and caught none of them. That is the honest line.
Before launch, on production, a decoy presented as a credential was refused and its agent suspended. That held directly, beside a real credential, through the sidecar, and with a suffix glued on.
Where the edge is
Tripwires catch an agent that uses or repeats a credential it found. They do not catch one that leaves the decoy alone, encodes it a way we don't claim, or carries it out a route PassControl cannot see.
One example, found in that run: PassControl Cloud runs on Vercel, whose edge strips the Proxy-Authorization header before any application sees it. A decoy hidden there is never checked. It also never authenticates anything, or reaches a provider.
If a trip was not a false alarm, rotate the agent's real key or passport before you resume it.
Try it
Tripwire keys are part of Behest v0.1 beta, on PassControl Cloud: passcontrol.vertias.eu. Plant them with the command above, then restart the sidecar so it checks outbound credentials.