Say you did the hard things. Every agent call goes through a gateway that checks which tool is called, who the message goes to and that each call carries a fresh, single-use proof. An agent can still be talked into trying anything, but out-of-scope calls are denied. What can an attacker still do?
That is the question behind Identity Shadow, the free browser chapter that finishes our Prompt Heist series. You play against a fictional hardened company across four acts. This post is the short version of what is left.
In-scope misuse: permitted does not mean intended
Enforcement keeps a call inside its limits. It does not ask whether this call, to this recipient, carrying this data, should happen. An agent allowed to message approved recipients can be persuaded to put a secret in that message. Act I is exactly that: the gateway denies everything out of scope, and the permitted send goes through.
The path around enforcement
A stolen token that is useless at the gateway can still work at an older endpoint that never went behind it. Single-use proofs and session keys protect only the routes that check them. Act II has a bearer token refused at the front door and accepted at a side door.
Identity confusion: a name is not an identity
If an approver matches a request by a display name, a second agent with the same name receives the approval. The real agent has a different identity that nobody compared. Act III shows an approval landing on the wrong agent.
Chained gaps
None of these breaks the gateway. A name earns an approval, an allowed recipient receives a report and a door that skipped enforcement accepts a credential found in it. Each step was permitted by the part of the system that checked it. Act IV chains them.
What closes each one
- Narrow what an agent can reach, what data it holds and where each action can send it. Keep secrets out of its context where you can, and ask a person to approve sensitive sends.
- Put every route to the data behind the same enforcement. Retire endpoints that accept a bare bearer token, and rotate anything that has appeared in a log.
- Bind approvals to a verified agent identity, not a display name, show the approver that identity and refuse a new agent that reuses a name in use.
Where Oathera fits
Oathera does not inspect or rewrite prompts, and it does not scan content for injected instructions. It gives every AI agent an identity bound to the machine it runs on, with a named person who answers for it, and checked every time it acts. Each call carries a single-use operation proof, and Oathera Agent Enforcement asks the Oathera Identity Service for a signed, single-use decision before the call goes through. That limits what a manipulated agent can do. It does not stop the manipulation, and it does not replace the checks inside your own systems or the work of finding paths that skip enforcement.
Play it, then see it refused
Identity Shadow is free, runs entirely in your browser and sends nothing you type anywhere. It unlocks after you finish Multi-Agent Conspiracy, the fifth chapter in the series. Play it at oathera.ai/identity-shadow. To watch a stolen token, a replayed proof and a tampered call get refused, try the simulator at sim.oathera.org. If you are putting agents into production, become a design partner.