Cognitive

Part 7 · Agentic AI Security

ASI07: Insecure Inter-Agent Communication

Agents that trust unauthenticated messages can be spoofed, misled, or silently redirected.

3 min readBuilds on Multi-Agent Systems

OWASP ASI07. This material is for defense. Only test systems you own or are authorized to test.

A forged message can spread through a shared communication channel.

What it is

Multi-agent systems depend on continuous communication through APIs, message buses, and shared memory. Weak authentication, integrity, confidentiality, or authorization lets attackers intercept, manipulate, spoof, replay, or block those messages.

How the attack unfolds

Without message authentication, an instruction from a stranger can look like one from a colleague.

Figure 46.1. Without message authentication, an instruction from a stranger can look like one from a colleague.

Common patterns

  • Message tampering in transit.
  • Agent spoofing through fake identities or descriptors.
  • Replay attacks that resend old valid messages.
  • Protocol downgrade to weaker communication modes.
  • Hidden instructions carried inside otherwise valid messages.

Defenses

Messages should be verified before an agent acts on them.
  1. Secure channels. Use encryption and mutual authentication.
  2. Sign messages. Protect the integrity of content and context.
  3. Prevent replay. Use nonces, session IDs, and timestamps.
  4. Pin protocol versions. Reject downgrades.
  5. Verify discovery. Authenticate agent registries and descriptors.
  6. Treat content as untrusted. A signature proves the sender, and not that the content is safe.

Common misconception

Common belief

"Our agents are internal, so they can trust each other."

In practice

A compromised internal agent can still send malicious messages.

Key takeaways

  • Agent messages can be spoofed, altered, or replayed.
  • Authenticate and sign communication between agents.
  • A verified sender can still deliver unsafe content.