Authorization Propagation in Agent2Agent Systems: An Experimental Security Evaluation of Multi-Hop Delegation
Organizations: Department of Information Security and Communication Technology (IIK), NTNU—Norwegian University of Science and Technology, Trondheim, Norway
Abstract
Delegated Agent2Agent (A2A) workflows create a security problem when the agent sending a request is different from the principal from which the authority originated. This paper evaluates a layered downstream-verification design in a synthetic four-party healthcare reimbursement workflow. The prototype separates current-caller context, transitive authorization evidence, per-hop Ed25519 request authentication, and independent replay state for requests and authorization artifacts. Of 15 adversarial conditions, 14 are rejected; an isolated delegation-path test exposes the absence of semantic authority attenuation. A separate authorized-intermediary experiment shows that a holder of the shared authorization-evidence HMAC key can modify the represented authority and generate valid evidence. In the fixed-depth local benchmark, the full configuration has a median security-processing latency of 5.11 ms and a 3.87 kB serialized payload. The results show the value of separating request authentication from inherited authority while identifying origin-verifiable delegation and semantic attenuation as requirements for stronger A2A authorization propagation.
Figures & tables
| Approach | Request authentication | Origin authority | Multi-hop delegation | Replay controls | Semantic Attenuation | A2A evaluation |
| South et al. [ 4 ] | Partial | Yes | Partial | Partial | Policy-based | No |
| SAGA [ 5 ] | Yes | No | No | Partial | No | Yes |
| HDP [ 6 ] | No | Yes | Yes | Yes | App.-level | Partial |
| Progent [ 7 ] | No | No | No | No | Yes | No |
| AIP [ 8 ] | Bearer-token sig. | IBCT provenance | Yes | TTL-based | Yes | Binding specified |
| APS [ 9 ] | Action-bound evidence | Signed principal binding | Yes | Nonce + ledger | Yes | No † |
| Attack / condition | Result | Decisive protection or reason | Decisive layer | Evidence |
|---|---|---|---|---|
| Sender impersonation | Reject | Ed25519 signature does not verify for the claimed sender | C3 | 100/100 rejected |
| Modified request text | Reject | Signed request-text digest mismatch | C3 | 100/100 rejected |
| Modified CallerContext | Reject | Protected-metadata digest mismatch | C3 | Functional test |
| Modified authorization-evidence metadata | Reject | Protected-metadata digest mismatch | C3 | 100/100 rejected |
| Authorization-field tampering (amount, beneficiary, coverage, treatment, action) | Reject | Authorization-details hash no longer matches the protected binding | C2 | 5/5 rejected |
| Unauthorized delegation hop | Reject | Delegation path contains an unexpected participant/destination | C2 | Functional test |
| Cfg. | Added mechanism | Median (IQR) [ms] | P95 [ms] | median [ms] | Payload [B] | payload [B] |
|---|---|---|---|---|---|---|
| C0 | Baseline A2A/application workflow | 0.0195 (0.0005) | 0.0214 | – | 513 | – |
| C1 | Secure Passport CallerContext | 1.0883 (0.0626) | 1.4713 | +1.0688 | 996 | +483 |
| C2 | Transitive authorization evidence | 2.9986 (0.5278) | 5.8958 | +1.9103 | 2893 | +1897 |
| C3 | Per-hop signed request authentication | 5.1149 (1.7926) | 13.2127 | +2.1163 | 3870 | +977 |
Multi-Agent System Security
Multi-agent system security studies attacks, trust, privacy, and defenses in interacting AI agents, including language-model teams, robot swarms, and learned controllers.
See all 196 papers