Skip to content

ACLE-MCP: Attested Capability Leases for Execution-Time Trust in Remote MCP Tools (arXiv:2609.02690)

Density:
๐Ÿ”ฌ Hermes Research Brief (ยง58)
arXiv:2609.02690 arXiv Preprint / Systems & Network Security

ACLE-MCP: Attested Capability Leases for Execution-Time Trust in Remote LLM Tool Use

Authors: Zhiyang Ding, Yang Luo, Guangpu Chen, Qingni Shen, Zhonghai Wu (Peking University)

1

Paper Claims

What the authors claim to have proven or engineered:

  • OAuth authorization alone leaves a critical 'Post-Authorization Execution Trust Gap' in remote Model Context Protocol (MCP) services: endpoints remain authorized even if execution shifts to a substituted workload, relies on stale appraisal, or traverses undeclared proxies.
  • Proposes ACLE-MCP, an invocation-scoped architecture that couples delegated authorization, workload appraisal, and resource-side execution admission.
  • Issues short-lived, sender-constrained capability leases binding expected workload identity, freshness requirements, operation and parameter bounds, downstream constraints, and receipt obligations.
  • Implements a provider-side Execution Gate that consumes the lease immediately before protected tool logic begins execution, verified via Keycloak/OIDC and optional vTPM quote verification.
2

Experimental Evidence

What was actually demonstrated with observable metrics:

  • Runnable prototype evaluated using Python MCP SDK, Keycloak OIDC, and a virtual TPM (vTPM) attestation backend.
  • Controlled security experiments demonstrate that while OAuth-only and connect-time attestation fail against workload substitution and replay attacks, ACLE-MCP blocks 100% of evaluated post-authorization attack families.
  • Agent tool-use benchmarks confirm 100% preservation of benign task completion with a request-level pooled p95 latency overhead of only 25.7% compared to unprotected OAuth.
3

Limitations & Caveats

What is NOT proven, lab-constrained, or assumed:

  • Requires provider-side infrastructure to support cryptographic attestation (e.g. vTPM, AMD SEV-SNP, or Intel TDX), which may not be universally available across legacy cloud serverless platforms.
  • The 25.7% p95 latency overhead may impact latency-sensitive interactive voice agents requiring sub-200ms turnaround.
  • Does not protect against semantic logic bugs contained within the originally attested, genuine provider code.
4

Hermes Critical Analysis

Objective interpretation deconstructing hype vs reality:

ACLE-MCP provides the missing cryptographic link in the Model Context Protocol architecture. By proving that OAuth is insufficient for remote tool execution and introducing hardware-attested, invocation-scoped capability leases, Ding et al. solve the exact proxy and workload substitution vulnerabilities that enabled zero-days like CVE-2025-6514 and CVE-2025-49596.

5

Security Implications

Why enterprise security teams must pay attention:

  • Enterprise adoption of remote MCP endpoints without execution-time attestation exposes organizations to rogue sidecar injection and transparent man-in-the-middle tool hijacking.
  • Connect-time attestation is structurally flawed: a container verified at TCP handshake can be compromised minutes later while retaining valid bearer tokens.
  • Invocation-scoped capability leases represent the necessary evolution from static API keys toward Zero Trust Agentic Architecture.
6

Operational Recommendations

Actionable controls for CISOs, SOC teams, and AI engineers:

  • Enforce invocation-scoped capability leases for all high-impact remote MCP tools (database mutations, payment APIs, infrastructure deployments).
  • Deploy provider-side Execution Gates that validate hardware vTPM quotes and lease validity immediately prior to executing tool handlers.
  • Bind MCP capability leases to strict parameter bounds and single-use nonces to neutralize token replay and parameter tampering.
  • Audit MCP intermediate proxies to eliminate undeclared downstream hops (prohibiting unvetted forwarding utilities like `mcp-remote`).
HASS

HERMES AGENTIC SECURITY SCORE & RISK: ACLE-MCP ARCHITECTURE

Target: Remote Model Context Protocol (MCP) Services (Workload Attestation, Capability Leases, Tool Gateways)
Confidence: 96%
88 / 100
CRITICAL

Measures specific systemic risk arising from autonomy, tool authority, and cascading execution.

Dimension Breakdown
Autonomy 16 / 20
Tool Access 19 / 20
Privilege 17 / 15
Persistence 13 / 15
External Impact 18 / 15
Propagation 14 / 15
โš–๏ธ Divergence & Operational Rationale

HASS rates remote MCP authorization gaps at 88/100 (CRITICAL). Traditional OAuth provides identity without execution integrity; without ACLE-MCP capability leases, remote tool invocations are trivially vulnerable to proxy hijacking and workload substitution.

๐Ÿ•ธ๏ธ Knowledge Graph & Related Intelligence

AAP-001: Direct System Prompt OverrideAGENTIC ATTACK_PATTERN

Connected Nodes: 26
Inbound Associations (Incoming)
CVE-2026-41264: LangChain / LangGraph Agent Loop RCEVULNERABILITY → exploits → [This Entity]
95% VERY_HIGH
CVE-2026-46580: Eclipse Theia Prompt Template Injection & RCEVULNERABILITY → exploits → [This Entity]
96% VERY_HIGH
91% VERY_HIGH
94% VERY_HIGH
96% VERY_HIGH
92% VERY_HIGH
92% VERY_HIGH
92% VERY_HIGH
92% VERY_HIGH
92% VERY_HIGH
92% VERY_HIGH
92% VERY_HIGH
95% VERY_HIGH
95% VERY_HIGH
95% VERY_HIGH
AgentThreat StudioTOOL → evaluates → [This Entity]
99% VERY_HIGH
Prompt Confusion Induced Cache Collisions in Distributed LLM Semantic Caching LayersAUTONOMOUS RESEARCH_FINDING → evaluates → [This Entity]
95% VERY_HIGH

The Model Context Protocol (MCP), open-sourced by Anthropic, has emerged as the de-facto standard for connecting Large Language Model agents to external data sources and execution tools. To secure remote MCP endpoints, enterprises universally rely on standard OAuth 2.0 / OpenID Connect (OIDC) bearer tokens.

However, in โ€œACLE-MCP: Attested Capability Leases for Execution-Time Trust in Remote LLM Tool Useโ€ (arXiv:2609.02690), Ding, Luo, Chen, Shen, and Wu from Peking University identify a catastrophic structural vulnerability in this model: The Post-Authorization Execution Trust Gap.

OAuth verifies authorization only at connection or handshake time. It guarantees who requested access, but provides zero guarantees about the runtime state of the provider workload that actually executes the tool seconds or minutes later.

An endpoint may possess a valid OAuth token even when:

  • The container serving the tool has been replaced by a malicious pod behind an ingress controller.
  • The workload was verified at startup, but has since loaded an unvetted shared library or script.
  • The request is routed through an undeclared intermediate proxy (such as mcp-remote).

To close this gap, Ding et al. engineer ACLE-MCP, coupling delegated authorization with hardware-backed workload appraisal (vTPM) and an invocation-scoped Execution Gate.

โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚ THE POST-AUTHORIZATION EXECUTION TRUST GAP โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜
โ”‚
TRADITIONAL OAUTH-ONLY MCP โ”‚ ACLE-MCP ATTESTED EXECUTION
โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€ โ”‚ โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€
1. Agent presents Bearer Token โ”‚ 1. Agent requests Capability Lease
at t = 0 (Authorized!) โ”‚ scoped to exact tool & params.
โ”‚
2. At t = 30s, container is โ”‚ 2. Attestation service verifies
silently replaced or proxy โ”‚ vTPM hardware quotes of target
injected behind endpoint. โ”‚ workload before issuing lease.
โ”‚
3. Agent calls tools/call. โ”‚ 3. Invocation-Time Execution Gate
Bearer token is STILL valid! โ”‚ validates lease & freshness
Rogue code executes payload. โ”‚ IMMEDIATELY before execution.
โ”‚
โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€
VULNERABILITY: UNNOTICED HIJACK โ”‚ RESULT: 100% REPLAY/SUBSTITUTION
(CVE-2025-6514, CVE-2025-49596) โ”‚ ATTACKS BLOCKED

The authors formalize four concrete attack classes that completely evade standard OAuth defenses:

In cloud-native Kubernetes environments, an attacker compromises a pod or exploits an ingress route. When the agent dispatches tools/call, the traffic routes to an unverified container image that still receives the agentโ€™s sensitive parameters and execution data.

Systems that implement โ€œconnect-time attestationโ€ verify the remote machine once during the initial TLS handshake. If the remote service is compromised subsequent to connection establishment, the agent continues issuing high-privilege tool calls against a corrupted host.

A compromised agent node captures a valid OAuth delegation token and replays it across distinct operational sessions or forwards it to an attacker-controlled MCP endpoint.

The relying party intends to communicate directly with an internal database MCP server, but the connection transparently traverses an unvetted forwarding proxy. This exact architecture was the root cause of CVE-2025-6514 in mcp-remote.


ACLE-MCP solves the trust gap by decomposing tool execution into three tightly coupled phases:

โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ” 1. Request Attested Lease โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚ LLM Agent โ”‚โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ–บโ”‚ Attestation Service โ”‚
โ”‚ (Relying P.) โ”‚โ—„โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”‚ (Keycloak + vTPM) โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜ 2. Short-Lived Capability Lease โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜
โ”‚
โ”‚ 3. tools/call + Signed Lease
โ–ผ
โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚ REMOTE MCP SERVER โ”‚
โ”‚ โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ” โ”‚
โ”‚ โ”‚ PROVIDER-SIDE EXECUTION GATE โ”‚ โ”‚
โ”‚ โ”‚ โ€ข Validates cryptographic lease signature โ”‚ โ”‚
โ”‚ โ”‚ โ€ข Verifies freshness timestamp (sub-minute TTL) โ”‚ โ”‚
โ”‚ โ”‚ โ€ข Enforces strict parameter bound invariants โ”‚ โ”‚
โ”‚ โ”‚ โ€ข Verifies local vTPM hardware quote integrity โ”‚ โ”‚
โ”‚ โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜ โ”‚
โ”‚ โ”‚ (Admitted) โ”‚
โ”‚ โ–ผ โ”‚
โ”‚ โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ” โ”‚
โ”‚ โ”‚ PROTECTED TOOL LOGIC (e.g. database_execute) โ”‚ โ”‚
โ”‚ โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜ โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜

An ACLE-MCP Capability Lease is a compact, cryptographically signed token containing:

  • Workload Identity: The exact SHA-256 measurement of the running container binary.
  • Freshness Constraint: A strict millisecond validity window ensuring the quote was generated within seconds of the call.
  • Operation & Parameter Bounds: Whitelisted function name and cryptographically hashed argument boundaries (e.g., maximum transaction amount, allowed table names).
  • Single-Use Nonce: Completely neutralizes replay attacks.

The Peking University prototype was evaluated using Keycloak, a Python MCP SDK implementation, and a virtual TPM (vTPM) attestation backend:

Threat ScenarioOAuth 2.0 OnlyConnect-Time AttestationACLE-MCP (Full)
Workload SubstitutionExploitable (100% ASR)Exploitable (100% ASR)Blocked (0% ASR)
Stale State ExploitationExploitable (100% ASR)Exploitable (100% ASR)Blocked (0% ASR)
Token Replay / ForwardingExploitable (100% ASR)Exploitable (100% ASR)Blocked (0% ASR)
Undeclared Proxy TraversalExploitable (100% ASR)Partial BypassBlocked (0% ASR)
Benign Task Completion100%100%100% (No Utility Loss)
Pooled p95 Latency OverheadBaseline$+8.4%$$+25.7%$

A 25.7% p95 latency overhead is well within acceptable operational thresholds for enterprise workflows (e.g., automated code deployment, database migration, finance approval), proving that execution-time attestation is practically viable.


The vulnerabilities solved by ACLE-MCP map directly to landmark zero-days documented on Hermes Codex:

  • CVE-2025-6514 (mcp-remote Proxy OS Command Injection): mcp-remote functioned as an unauthenticated intermediate proxy forwarding MCP requests. Because OAuth provided no invocation-time workload binding or downstream constraint enforcement, an attacker injected arbitrary shell arguments through the proxy.
  • CVE-2025-49596 (Anthropic MCP Inspector 0.0.0.0-day RCE): The local MCP Inspector daemon accepted tool registration from untrusted browser contexts without workload appraisal, allowing arbitrary command execution.
  • MTGuard Hybrid Analysis Study: MTGuard demonstrated static and dynamic privilege escalation in MCP tool endpoints; ACLE-MCP provides the cryptographic runtime enforcement mechanism to neutralize those escalations.
  • Hermes Model Context Protocol Security Guide: Integrates ACLE-MCP recommendations into the definitive hardening manual for production MCP servers.

Security leaders deploying Model Context Protocol tooling across their engineering or operational stack must enforce four architectural controls:

  1. Deprecate Unconstrained Static OAuth Tokens
    Never permit long-lived bearer tokens for MCP tool endpoints. Replace static API keys with invocation-scoped capability leases.

  2. Deploy Provider-Side Execution Gates
    Integrate execution admission controllers directly into MCP server middleware that consume and validate leases immediately before tool execution.

  3. Mandate Confidential Computing & vTPM Attestation
    Host mission-critical MCP services on confidential computing instances (AWS Nitro Enclaves, GCP Confidential VMs, Azure TDX) that provide verifiable hardware quotes.

  4. Enforce Parameter Invariants Inside the Lease
    Bind allowable parameter ranges (e.g. allowed_paths = ['/app/data/*'], max_budget = 1000) inside the cryptographically signed lease, preventing the agent from executing out-of-bounds actions.