-
Notifications
You must be signed in to change notification settings - Fork 2
research(security): AIP — Invocation-Bound Capability Tokens for verifiable agent delegation across MCP and A2A #2504
Description
Source
arXiv:2603.24775 — "AIP: Agent Identity Protocol for Verifiable Delegation Across MCP and A2A" (March 25, 2026)
Finding
AIP introduces Invocation-Bound Capability Tokens (IBCTs): short-lived JWT/Biscuit tokens that bind:
- Caller identity (agent ID + signing key)
- Delegated capability (specific tool + argument constraints)
- Invocation scope (single-use or session-scoped)
A scan of ~2,000 public MCP servers found all lacked authentication — any client can call any tool without identity verification. A2A also lacks a standard delegation mechanism: agents can impersonate each other's tool calls.
Applicability to Zeph
zeph-mcp and zeph-a2a have no identity propagation:
- MCP: tool calls arrive without caller identity — trust level assigned per-server, not per-caller
- A2A: agent invocations carry no verifiable delegation token
Note: #2497 (SMCP hardening) and #2496 (threat modeling) address tool poisoning and shadowing attacks but neither covers the identity/delegation gap this paper targets.
Implementation sketch:
- Add
caller_token: Option<Ibct>toToolCallinzeph-mcp McpManager: validate IBCT signature + scope before routing tool invocations- A2A client: attach delegation token when invoking sub-agents; A2A server: verify token
- Config:
[mcp.identity] require_tokens = false(off by default for backward compat with existing servers)
Priority rationale
P2: the scan result (0/2000 servers authenticate callers) confirms this is a systemic gap. The IBCT mechanism is lightweight enough to implement incrementally without breaking existing server compatibility.