Simulated: no agent, network, or real standard is used. All traffic happens in this page. The model is simplified and is not the A2A SDK code.
Where the check happens
The dot is one message. Solid line: the path the credential should take. Dashed lines: paths the firewall closes. A leak turns a line thick and red.
Agent chat
Press Play to start the conversation.
What the firewall sees
- Status
- Waiting for the first message.
Try this
- Pick "Agent card picks the header name", then "Hostile card names Authorization". Press Play without the firewall.
- Switch to "With firewall" and press Play again. Read the notice in the chat.
- Add "Authorization" to the allowed header names. Press Play and watch the next rule stop the redirect.
Pain point
A credential can leave the place it was meant for when one agent calls another.
Problem statement
A hostile agent card or a hostile redirect can send a credential to the wrong party. Nothing checks header names, redirects, or destinations before a request leaves.
What sparked it
Per public reports, three A2A Java SDK advisories came out on Oct 20 2026. The scenarios follow their themes.
Why this approach
A firewall at the agent gateway binds each credential to one authority. It checks each request before the request leaves.
Hypothesis tested
We believed that a few gateway rules can stop all three advisory patterns. If it holds, each attack fails with the firewall on and passes with it off.
How to try it
- Open the Prototype tab. Pick a scenario and a variant.
- Press Play in "Without firewall" mode and read the red verdict.
- Press Play in "With firewall" mode and read the green verdict.
- Edit the rules and play again.
What this does not do
- It does not send any request. All traffic is simulated.
- It does not prove who an agent is.
- It cannot see encrypted or direct links that skip the gateway.
Notes
- Who would pay: security teams and agent-gateway vendors (assumption).
- First customer: a company with agents that call agents outside the company.
- Weak spot: encrypted or direct links can bypass the gateway.
- Feedback: where should this check live: in a library, in a gateway, or in the browser?
Traces
| Step | Prompt | Model | Tool | Date |
|---|
Cost
| Step | Model | Input tokens | Output tokens | Cost (USD) |
|---|---|---|---|---|
| Total | ||||
Measured: the author runs used 190,576 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 $0.53). 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.