Identity · CA
A Security Scan Found Every MCP Server It Tested Had No Authentication
A new paper proposing an Agent Identity Protocol (AIP) for verifiable delegation across MCP and A2A opens with a blunt diagnosis of why agent identity is still unsolved, not just under-specified.
The authors note that the two dominant protocols for giving agents the ability to act — Anthropic's Model Context Protocol for tool invocation and Google's Agent-to-Agent protocol for inter-agent collaboration — define how agents communicate but not who they are. MCP added OAuth 2.1 as an optional authorization layer in 2026, but a security scan by Knostic of roughly 2,000 live MCP servers found that every single one lacked authentication. On the A2A side, agent cards carry self-declared identities with no attestation binding them to anything real. The practical consequence: when one agent delegates a task to another, nothing verifies the delegating agent's authority, nothing constrains the receiving agent's scope, and nothing records the handoff for later audit.
The paper surveys the current patchwork of attempted fixes — Google DeepMind's macaroon-based Delegation Capability Tokens, four IETF Internet-Drafts published in early 2026 covering pieces like agentic JWTs and SCIM for agents, and Mastercard's Verifiable Intent proposal for commerce — and argues none of them actually ships a working protocol that handles offline delegation, chained policy, and provenance records across both MCP and A2A at once. It also flags structural limits in existing building blocks: OAuth 2.1 needs a centralized authorization server and produces opaque tokens with no visible delegation chain, while W3C DIDs bring blockchain dependencies and circular trust bootstrapping that make them awkward for this specific job.
What's notable isn't that someone proposed yet another identity scheme — plenty of those exist. It's the empirical detail underneath the proposal: a near-total absence of authentication across thousands of deployed MCP servers in production use. That's not a theoretical gap discovered in a lab; it's the current state of a protocol that's already being wired into enterprise tool-calling pipelines. Any orchestration layer built on top of MCP today is, by this measurement, delegating work through servers with no way to verify who's asking.
The fix the paper proposes borrows attenuation-first design from macaroons rather than OAuth's centralized model, aiming for delegation chains that can be checked offline and audited after the fact — closer to attestation infrastructure than a login system. Whether AIP itself gets adopted matters less than the diagnosis: agent-to-agent delegation at scale is running ahead of any mechanism to verify who delegated what to whom.