Post-Authorization Execution Trust Gap in Remote MCP Tool Invocation
First seen Sep 3, 2026 · Updated Sep 3, 2026
This is a defensive research paper, not an active exploit, that identifies a structural weakness in OAuth-secured remote MCP tool use: a tool endpoint can remain 'authorized' even after the underlying workload executing the call has been substituted, is running stale attestation state, or reuses authority meant for a different sender. The authors propose ACLE-MCP, an invocation-scoped capability-lease architecture that binds authorization to fresh, workload-specific execution state, and demonstrate it closes these gaps with a manageable latency cost.
Technical Analysis
The core issue is that standard OAuth-based authorization for MCP tool calls validates identity/permissions at connect or token-issuance time but does not re-verify, at the moment of execution, that the same attested workload is still the one performing the protected operation. This creates several concrete attack classes: workload substitution after authorization, reliance on stale appraisal/attestation state, reuse of delegated authority by an unintended downstream sender, and silent traversal through undeclared downstream components (transitive tool chaining). An attacker or compromised infrastructure component could exploit this gap to execute tool logic under a previously valid authorization even though the actual execution environment has changed, effectively a confused-deputy/TOCTOU-style problem at the protocol layer. ACLE-MCP mitigates this by issuing short-lived, sender-constrained capability leases bound to workload identity, freshness, operation/object/parameter bounds, and downstream constraints, enforced by a provider-side Execution Gate immediately before tool logic runs, with optional vTPM-based quote verification for hardware-rooted workload attestation.
Affected Systems
MCP Python SDK, Keycloak/OIDC-based MCP servers; protocols: MCP, OAuth
Detection Signatures
- Log pattern: tool invocation timestamps significantly later than OAuth token issuance without re-attestation checkpoint
- Log pattern: mismatched workload identity/attestation quote between authorization time and execution time
- Log pattern: authorized calls traversing undeclared or unexpected downstream MCP servers/components
- Log pattern: reuse of a delegated token/authority across a different sender identity than originally granted
- Absence of invocation-time freshness or nonce validation in MCP server tool-call handlers
Remediation Steps
- 1
Adopt invocation-scoped authorization
Do not rely solely on connect-time or token-issuance-time OAuth checks; bind authorization to the specific execution instance via short-lived, sender-constrained capability leases as proposed in ACLE-MCP.
- 2
Enforce execution-time workload attestation
Deploy a provider-side Execution Gate that verifies current workload identity and freshness (e.g., via vTPM quotes) immediately before executing protected tool logic, not just at connection setup.
- 3
Constrain and log downstream tool chaining
Explicitly declare and validate allowed downstream components/tools in the capability lease so that authority cannot be silently extended through undeclared multi-hop tool calls.
- 4
Set short authorization freshness windows
Reduce the validity window of tool-call authorizations and require re-validation for long-running or resumed sessions to limit exposure from stale appraisal state.
- 5
Benchmark latency trade-offs before production rollout
Evaluate the ~25% p95 latency overhead reported in the paper against operational SLAs before adopting full invocation-time binding in latency-sensitive agent pipelines.
Industries Most Exposed
Respond to this threat
Pro subscribers get a full AI-generated incident-response playbook for this threat — detection, containment, eradication, and recovery steps — plus an unlimited AI Threat Advisor for questions about your environment.