Security teams are starting to understand which AI tools employees use and which MCP servers those tools connect to. Far fewer can say which agents hand work off to other agents. The A2A protocol makes that kind of delegation easy to set up and hard to detect, which is how blind spots usually form.
Why A2A Communication Falls Outside Existing Security Tools
The A2A protocol moved quickly. Google introduced it in April 2025, and by June it was a Linux Foundation project backed by more than 100 technology companies. Most security tooling was built for people using apps and for apps calling APIs, and A2A traffic looks like the latter but behaves more like the former. A gateway sees a well-formed HTTPS request from an authenticated service. It does not see that the body is a task written by another model.
The protocol is also opaque by design. The specification says agents collaborate “without needing access to each other’s internal state, memory, or tools,” so the calling agent sees only the results, not the steps the remote agent took to produce them. Multi-agent systems built within one platform face the same problem. Pluto’s research on Copilot Studio found that a connected subagent “exists only as a piece of the parent’s configuration,” with no identity of its own to log against.
How the A2A Protocol Widens the Attack Surface Beyond MCP
A2A vs. MCP comes down to who is on the other end. MCP connects an agent to its tools and data. A2A connects agents to each other, and the remote side is a reasoning system that can call its own tools and other agents, so one request may fan out across teams or companies.
MCP alone is already hard to track. Pluto’s MCP Debrief found that the median organization runs 288 distinct MCP servers, and 82% of installations run on servers that ingest text written by someone else. A2A adds another layer: untrusted text arrives as a task from a peer instead of a tool result.
Discovery adds its own exposure. An agent advertises its skills, endpoint, and authentication requirements in an agent card, usually served at /.well-known/agent-card.json. The spec says cards “MAY be digitally signed,” making signing optional; an unsigned card is an unverified claim about what an agent is.
The Trust Gap When Agents Accept Tasks From Other Agents
A2A supports standard authentication for the connection but leaves the decision to the receiver. The spec states that “authorization boundaries are defined by each agent’s authorization model, not prescribed by the protocol.” Whether a request should be honored depends entirely on whoever built the receiving agent.
A common mistake is treating an authenticated peer as trusted. Authentication proves which agent sent a task. It does not prove the task reflects what any person wanted. If the sending agent was steered by a poisoned document earlier in its session, its next request arrives with valid credentials and a clean-looking payload.
Delegated authority makes this worse. A receiving agent running with broad service credentials can end up doing for a low-privilege caller what that caller could never do directly: the classic confused deputy problem with a new transport.
What Securing A2A Communication Requires
OWASP’s Top 10 for Agentic Applications now lists insecure inter-agent communication as its own risk, ASI07. Closing it takes controls on both sides of each connection.
- Inventory: Know which agents publish an agent card and which agents call them, including ones built in low-code platforms.
- Verified identity: Use mutual authentication and require signed agent cards, even though the spec leaves signing optional.
- Per-task authorization: Check whether this caller may request this skill on behalf of this user, instead of trusting any authenticated peer.
- Untrusted task content: Validate and screen incoming tasks the way you would any external input.
Logging must follow the chain, too. An audit trail that stops at the first agent cannot show where a bad instruction actually started.
Final Thoughts
A2A solves a real interoperability problem, and its spec does cover authentication and input validation. It still leaves the hardest calls, who to trust and what to allow, to whoever builds each agent. Until security teams can see agent-to-agent traffic and attach identity and policy to it, A2A security will trail adoption, and the blind spot will grow with every agent that learns to delegate.