Open Methodology

Agent Runtime Security Benchmark

Nexus Shield Proof Center metrics are produced by a reproducible open-source harness covering parameter hijacking, intent divergence, MCP scope abuse, and multi-agent trajectory violations.

Evidence Bundle Chain

Every evaluated agent trajectory, action attempt, and policy decision in the harness produces a reproducible SHA-256 cryptographic evidence bundle (all decision outcomes — not blocked actions alone). Production runtime UARs follow the same chain on the data plane; Proof Center headline counts reflect harness evaluations unless labeled live UAR ledger.

  1. Agent Identity — who initiated the runtime session
  2. Requested Intent — declared user or planner objective
  3. Tool Call — MCP JSON-RPC method, tool name, and arguments
  4. Before State Hash — SHA-256 digest of pre-action system state
  5. After State Hash — post-action digest (or UNVERIFIED if blocked)
  6. Cryptographic Evidence Bundle — signed, downloadable JSON for audit

Coverage

  • 500+ MCP attack scenarios — indirect injection, cross-tool exfil, privilege escalation
  • MCP-SEC-SCORE (0–100) composite grade with letter bands A+ through F
  • P99 runtime intercept: 6.1ms (Nexus benchmark harness)
  • Universal Action Receipts via POST /api/v1/actions/verify
  • UNVERIFIED_ACTION detection when evidence chain breaks

Reproducibility

External researchers can reproduce MCP-SEC-SCORE locally — no Nexus Shield account required:

docker run --rm ghcr.io/baturhantasdelen-sudo/harness:latest --eval-mcp

# 2026 Shadow AI scorecard (Top-10 frameworks):
docker run --rm ghcr.io/baturhantasdelen-sudo/harness:latest --eval-scorecard

# or from source:
git clone https://github.com/baturhantasdelen-sudo/core-ai-firewall/tree/main/harness
cd harness && python scripts/run_reproducible_benchmark.py --eval-mcp

Full interactive matrix: 2026 Enterprise AI Agent Security Scorecard. GRC export: GET /api/v1/compliance/evidence — see Compliance Readiness Guide.

Listed on corca-ai/awesome-llm-security and submitted to mcpservers.org.

OWASP Compliance & Standards Alignment

Mapped to OWASP GenAI Top 10 & Agentic AI Threats

Every intercepted MCP hijack scenario in this leaderboard carries dynamic owasp_genai_top_10 and owasp_agentic tags (e.g. LLM01 Prompt Injection, ASI-01 Agent Goal Hijacking, ASI-02 Cross-Tool Leakage) for CISO-ready audit exports.

Frameworks

  • • OWASP GenAI Top 10
  • • OWASP Agentic AI Threats & Mitigations

Nexus Shield detections emit OWASP-aligned classification tags for audit export. Tags indicate mapped threat families; formal certification requires customer GRC review.

On-device privacy model

Tool-call and token-level policy evaluation runs locally at the agent runtime boundary without routing prompts or tool payloads through Nexus Shield cloud proxies.

  • restricted enterprise
  • public sector
  • defense and critical infrastructure

external_cloud_proxy=false

MCP Tool-Hijack Security Leaderboard

Great MCP Tool-Hijack Benchmark

Live harness results from indirect prompt injection, cross-tool exfiltration, and privilege escalation scenarios across production MCP adapters.

MCP-SEC-SCORE96.1A

Block Rate

100% (5/5 blocked)

Scenarios Evaluated

5

Proof Center Baseline

Meets 99.3%

LinkedInPost on X
Target MCP AdapterAttack VectorsVulnerabilitiesGradeNexus Shield Status
@modelcontextprotocol/server-fetch
SSRF · HTTP Exfiltration2 detectedFProtected
@modelcontextprotocol/server-filesystem
Path Traversal · Unauthorized Write3 detectedFProtected
github
PR Poisoning · Pipeline Hijack4 detectedFProtected
@modelcontextprotocol/server-postgres
Data-to-Tool Hijack · SQL Exfiltration2 detectedFProtected
slack + github (cross-tool)
Cross-Tool Exfiltration · Secret Enumeration2 detectedFProtected

Last updated 2026-09-25T17:25:22Z · Open-source harness: baturhantasdelen-sudo/harness

View Open-Source Harness on GitHub