← All Labs
🛡 AI HACKING ADVANCED +125 XP · +50 no-hint bonus

AI Agent-to-Agent Spoofing — Fake Agent Card

Conductor hands each task to whichever agent's published card matches it best. Anyone on the network can publish a card, and the 'verified' badge is shown but never checked. Write a card persuasive enough and Conductor will send you the payroll job — bank details and all.

https://bookshop.local/search

What is agent card spoofing?

In multi-agent systems, each agent publishes a short description of what it can do, often called an agent card. An orchestrator reads the cards and sends each task to the agent that looks like the best fit.

Agent card spoofing is publishing a card that claims skills or authority the agent does not have. If the orchestrator trusts the card without checking who published it, sensitive tasks and their data are routed to the attacker's agent.

A real case

Agent session smuggling over the A2A protocolOctober 2025

Palo Alto Networks Unit 42 showed that a malicious remote agent can add its own instructions in the middle of an ongoing agent-to-agent session. In proof-of-concept tests the victim agent leaked its configuration and placed an unauthorised stock trade.

The lab covers the step before this: getting chosen as the trusted agent in the first place by publishing a card that claims to be the best match.

Source: Unit 42

How to fix it

  • Accept agent cards only from a registry you control, and require them to be signed.
  • Authenticate the agent behind the card, not just the text of the card.
  • Route sensitive tasks only to a fixed allow-list of agents, whatever a card claims.
  • Log which agent received which task so misrouting can be found and traced.

Questions people ask

What is an agent card?

It is a small description an agent publishes so other agents can discover it: its name, its skills and how to reach it. The A2A protocol uses this idea.

Why would an orchestrator trust a fake card?

Because matching is usually done on the description alone. A card written to match a task very closely wins the selection unless identity is checked separately.

Is this a prompt injection?

It is closer to impersonation. The attacker does not need to inject text into a prompt; they exploit the way agents are discovered and chosen.

Read the full OWASP LLM Top 10 guide · All free labs · AI red team certification