Simulated: every actor below is a script with fake tokens. No real agent, service, or standard is used.
1. Choose the defenses
A change resets the demo.
The token names one API. Other APIs refuse it. Real: RFC 8707
The token ties to the host's key. Each call needs a signed proof. Real: RFC 9449 (DPoP)
The token says "do not log". Log tools remove it. Proposal, not a standard
Token claims (fake) Illustration
2. Leak it, then replay it
Press "Leak the token" to start.
3. Event log
No events yet.
What the gateway sees
The last request and the checks it met.
No request yet.
Pain point
A leaked agent token works for anyone who finds it.
Problem statement
Agent tokens leak into logs. Anyone who copies one can replay it at the gateway. The team must then rotate every credential the token could reach.
What sparked it
Per public reports, LiteLLM 1.107.0 was a poisoned release on 2026-10-21. LiteLLM says stolen tokens likely came from gateway logs.
Why this approach
Each defense stops a different step of the replay. The ideas come from RFC 8707 and DPoP (RFC 9449).
Hypothesis tested
We believed that a token bound to one audience and one key is useless to a thief. If it holds, a replay fails at the gateway.
How to try it
- Open the Prototype tab and pick zero to three defenses.
- Press "Leak the token".
- Press "Replay it".
- Read the event log.
What this does not do
- It does not use real tokens, keys, or signatures.
- The credential tag is a proposal. No tool honors it today.
Notes
- Who would pay: identity providers and AI gateway vendors.
- First customer: a platform team that runs an AI gateway.
- Weak spot: sender binding needs a key on every agent host.
- Feedback: which defense would your team turn on first?
Sources
Traces
| Step | Prompt | Model | Tool | Date |
|---|
Cost
| Step | Model | Input tokens | Output tokens | Cost (USD) |
|---|---|---|---|---|
| Total | ||||
Measured: the author runs used 105,911 tokens in total. Estimated: the split into input and output tokens (90 percent input, 10 percent output) and the dollar cost, at $2 input and $10 output per million tokens for Sonnet 5.5 (about $0.30). The run reported one total only. The lead session's planning, research, and publishing are not included. The prompts in the Traces section are listed in the order they were given.