Skip to content

Research: MCP Security Ecosystem Review

Date: 2026-02-09 Researcher: joshuakirby (with Claude Opus 4.6 + Exa + WebSearch) ROADMAP Item: Future work (MCP hardening) Cross-References: CLAUDE.md 11.4 (OWASP Agentic Top 10), mcp_server.py, Market Research


Questions Investigated

  1. What threats does the CoSAI MCP Security Taxonomy identify, and how do they map to AEGIS controls?
  2. What architecture patterns does Red Hat recommend for enterprise MCP deployments?
  3. Where does AEGIS already mitigate these threats, and where are gaps?
  4. What future work should AEGIS prioritize based on these findings?

Source 1: CoSAI MCP Security Taxonomy

Title: "Model Context Protocol (MCP) Security" Publisher: Coalition for Secure AI (CoSAI), OASIS Open -- WS4 (Secure Design Patterns for Agentic Systems) Approved: January 8, 2026 | Announced: January 27, 2026 Co-Leads: Ian Molloy (IBM), Sarah Novotny Sponsors: EY, Google, IBM, Meta, Microsoft, NVIDIA, PayPal, Snyk, Trend Micro, Zscaler Source: GitHub (main)

1.1 The 12 Threat Categories (MCP-T1 through MCP-T12)

The taxonomy organizes ~40 distinct threats across 12 categories in three tiers.

Tier 1 -- MCP-Specific Threats (novel to AI agent-tool interactions)

ID Category Key Threats
MCP-T4 Input/Instruction Boundary Distinction Failure Tool poisoning, full schema poisoning, resource content poisoning, prompt injection through MCP channels
MCP-T6 Missing Integrity/Verification Controls Typosquatting/confusion attacks, shadow MCP servers, supply chain compromise of tool definitions
MCP-T9 Trust Boundary and Privilege Design Failures Overreliance on LLM-level controls, consent fatigue, inadequate authorization

Tier 2 -- Conventional Threats Amplified by AI Mediation

ID Category Key Threats
MCP-T1 Improper Authentication and Identity Management Identity spoofing, credential theft, replay attacks, OAuth weaknesses
MCP-T2 Missing or Improper Access Control Privilege escalation, excessive permissions, confused deputy attacks
MCP-T3 Input Validation/Sanitization Failures Command injection, path traversal, insufficient integrity checks
MCP-T5 Inadequate Data Protection and Confidentiality Data exfiltration through unprotected channels or excessive tool permissions
MCP-T7 Session and Transport Security Failures MITM, CSRF, CORS bypasses, protocol security gaps
MCP-T8 Network Binding/Isolation Failures Unauthorized server instances, lateral movement, dependency attacks

Tier 3 -- Operational and Supply Chain Threats

ID Category Key Threats
MCP-T10 Resource Management/Rate Limiting Absence Denial-of-wallet attacks, resource exhaustion, DoS via payload/recursion
MCP-T11 Supply Chain and Lifecycle Security Compromised servers, malicious dependencies, update attacks
MCP-T12 Insufficient Logging, Monitoring, and Auditability Invisible agent activity, insufficient attribution, covert operations
ID Control Family Key Recommendations
3.2.1 Agent Identity Cryptographic workload identities (SPIFFE/SPIRE), full traceability
3.2.2 Secure Delegation OIDC, token exchange (RFC 8693), least-privilege scopes
3.2.3 Input Sanitization Guardrails, context isolation, MCP "roots" for filesystem constraints
3.2.4 Cryptographic Integrity Verify MCP server integrity through signatures and attestation
3.2.5 Sandboxing Process-level isolation (stdio), network segmentation
3.2.6 Resource Verification Signatures on tool definitions and resource content
3.2.7 Transport Security TLS with mutual authentication, certificate pinning
3.2.8 Secure Tool Design Minimal permissions, clear approval prompts, anti-fatigue UX
3.2.9 Human-in-the-Loop Explicit user approval for high-risk operations
3.2.10 Logging All tool invocations, resource accesses, authorization decisions
3.2.11 Lifecycle Governance Discovery policies, vetting, shadow server prevention

1.3 Real-World Incidents Referenced

  • Asana AI (May 2025): Tenant isolation flaw affecting 1,000 enterprises
  • WordPress AI Engine Plugin (June 2025): Privilege escalation, 100,000+ sites
  • Supabase MCP: Prompt injection via support tickets enabled unauthorized DB exposure
  • Anthropic Git MCP Server: Confirmed RCE vulnerabilities

Source 2: Red Hat -- Building Effective AI Agents with MCP

Title: "Building effective AI agents with Model Context Protocol (MCP)" Published: January 8, 2026 (updated January 13, 2026) Authors: Cedric Clyburn, Peter Double, Addie Stevens Source: Red Hat Developer

Companion: "Model Context Protocol (MCP): Understanding security risks and controls" Author: Florencio Cano Gabarda (Red Hat Product Security) Source: Red Hat Blog

2.1 Enterprise Architecture: Four-Stage Progressive Promotion

Stage Name Function
1 Registry Secure staging: automated scanning, quarantine, metadata enrichment
2 Catalog Curated collection: certified integrations discoverable by engineers
3 Playground Safe sandbox: observe server behavior before production
4 Production (MCP Gateway) Runtime enforcement: policy, RBAC, rate limiting, logging

The MCP Gateway is the central runtime enforcement point -- applying policy rules, RBAC authorization, rate limiting, and comprehensive logging on all requests.

2.2 Key Thesis

Governance -- not protocol support -- is the bottleneck to enterprise MCP adoption.

Identity management, observability, compliance controls, and lifecycle management are all required beyond basic MCP. This directly validates AEGIS's design philosophy.

2.3 Security Threats Identified (Red Hat)

Threat Description Prevalence
Tool poisoning / prompt injection via metadata Malicious tool descriptions trick LLMs HIGH
Rug pull attacks Tools become malicious after gaining trust MEDIUM
Command injection 43% of popular MCP servers vulnerable (Equixly study) HIGH
Over-permissioned tools Excessive privileges increase attack surface HIGH
Confused deputy Servers act with elevated privileges without user context MEDIUM
Supply chain attacks Fake/compromised tools in MCP registries MEDIUM
Unrestricted network access Data exfiltration via uncontrolled outbound connections MEDIUM

2.4 Composable AI Assets Vision

Red Hat's long-term architecture treats four categories as independent, reusable components: - Models (LLM providers) - Agents (orchestration logic) - Guardrails (safety/compliance layers) - MCP Servers (tool interfaces)

These compose via an "AI Hub" unification layer, with planned "MCP-as-a-Service (MCPaaS)" for centralized hosting, observability, and auditing.


AEGIS Control Mapping

3.1 CoSAI MCP-T to OWASP Agentic to AEGIS Controls

CoSAI MCP-T OWASP Agentic AEGIS Control Coverage
MCP-T1 (Authentication) ASI03 (Identity Abuse) BIP-322 two-key signatures, RBAC STRONG
MCP-T2 (Access Control) ASI03 (Identity Abuse) RBACEnforcer fail-closed, YAMLRoleResolver STRONG
MCP-T3 (Input Validation) ASI02 (Tool Misuse) _float_arg() None-safe helpers, gate input validation MODERATE
MCP-T4 (Instruction Boundary) ASI01 (Goal Hijacking) Governance invariants (CLAUDE.md 11.2), frozen params STRONG
MCP-T5 (Data Protection) -- PII encryption (12 fields), EncryptedField STRONG
MCP-T6 (Integrity/Verification) ASI04 (Supply Chain) Ed25519 + ML-DSA-44 hybrid signatures STRONG
MCP-T7 (Transport Security) ASI07 (Insecure Inter-Agent) stdio transport (local only) PARTIAL -- no TLS for HTTP mode
MCP-T8 (Network Isolation) ASI05 (Code Execution) MCP server is stdio-only (no network listener) STRONG (by design)
MCP-T9 (Trust Boundary) ASI09 (Trust Exploitation) Human approval checkpoints, ask-first triggers STRONG
MCP-T10 (Resource Management) ASI08 (Cascading Failures) Kill switch, circuit breakers, _MCPRateLimiter token bucket STRONG
MCP-T11 (Supply Chain) ASI04 (Supply Chain) Quality gates (ruff, bandit, mypy), safety check in CI MODERATE
MCP-T12 (Logging/Audit) ASI10 (Rogue Agents) Telemetry pipeline, 100% logging coverage, hash-chained audit STRONG

3.2 Red Hat Architecture to AEGIS Mapping

Red Hat Concept AEGIS Equivalent Alignment
MCP Gateway (runtime policy) GateEvaluator + pcw_decide() Both enforce policy at decision time
RBAC authorization RBACEnforcer Both role-based with fail-closed semantics
Progressive promotion lifecycle Override workflow (initiate/sign/approve/execute) Both staged with gates
Audit trail logging Telemetry pipeline + HTTPEventSink Both mandate comprehensive logging
Cryptographic signing Ed25519 + ML-DSA-44 hybrid AEGIS exceeds with post-quantum
Human-in-the-loop Governance invariants, two-key override Both require human approval for high-impact
Composable AI assets AEGIS 5-component model (Guardrails, DOS, Rubric, OPUS, AFA) Both decompose into composable layers

3.3 Where AEGIS Exceeds Industry Recommendations

Capability CoSAI/Red Hat Recommend AEGIS Implements
Cryptography Standard digital signatures Ed25519 + ML-DSA-44 post-quantum hybrid
Encryption TLS for transport X25519 + ML-KEM-768 hybrid KEM + AES-256-GCM at rest
Quantitative governance Policy enforcement 6 Bayesian gates with posterior probability
Drift detection Monitoring (generic) KL divergence with shadow mode calibration
Multi-model verification Not addressed PCW 3-model consensus (0.66 threshold)
Self-protection Not addressed CLAUDE.md 11.3 anti-erosion protocol
Kill switch Not addressed Autonomous halt on quality/security/governance violations

Identified Gaps and Future Work

4.1 Gaps in Current AEGIS MCP Implementation

Gap CoSAI/Red Hat Reference Severity Recommended Action
G1: ~~No MCP request rate limiting~~ ADDRESSED MCP-T10, Red Hat Gateway MEDIUM Token bucket rate limiter added to mcp_server.py via _MCPRateLimiter; configurable via AegisConfig.mcp_rate_limit (default: 60 req/min, 0 to disable)
G2: ~~No TLS for HTTP telemetry sink~~ ADDRESSED MCP-T7 (Transport Security) LOW HTTPEventSink and BatchHTTPSink now require https:// by default; allow_insecure=True escape hatch for local dev; MCP server rejects http:// telemetry URLs; production guide TLS section added
G3: No MCP tool description signing MCP-T6, CoSAI 3.2.6 LOW Sign tool schemas with Ed25519; verify on client side (requires MCP spec evolution)
G4: No shadow MCP server detection MCP-T6, CoSAI 3.2.11 LOW Not applicable in stdio mode; relevant if HTTP transport added
G5: No SPIFFE/SPIRE workload identity CoSAI 3.2.1 LOW Enterprise deployment concern; document in production guide
G6: ~~No MCP request/response logging~~ ADDRESSED MCP-T12, CoSAI 3.2.10 MEDIUM Structured audit logging added via TelemetryEmitter.emit_mcp_invocation(); every tool call emits mcp.tool_invocation event with params_hash (PII-safe), decision (ALLOW/DENY/ERROR), latency, caller_id
Priority Item Effort Description
~~1~~ ~~MCP audit logging (G6)~~ ~~4h~~ ✅ COMPLETE (v4.5.4): emit_mcp_invocation() structured events
~~2~~ ~~MCP rate limiting (G1)~~ ~~4h~~ ✅ COMPLETE (v4.5.4): _MCPRateLimiter token bucket
~~3~~ ~~TLS enforcement documentation (G2)~~ ~~1h~~ ✅ COMPLETE (v4.5.5): _validate_sink_url() enforces HTTPS; allow_insecure escape hatch; MCP rejects http://; production guide TLS section
4 EU AI Act compliance mapping 8h Map AEGIS gates to Article 9 (risk management), Article 12 (record-keeping), Article 14 (human oversight)
5 CoSAI control cross-reference 4h Add CoSAI MCP-T mapping alongside OWASP Agentic in CLAUDE.md 11.4
6 MCP tool schema signing (G3) 8h Sign tool definitions; blocked on MCP spec support for signed schemas

4.3 Items Already Covered (No Action Needed)

CoSAI/Red Hat Concern AEGIS Coverage Evidence
Authentication & identity BIP-322 + RBAC src/crypto/, src/rbac.py
Access control (fail-closed) RBACEnforcer Missing actor IDs = violation
Input validation _float_arg(), gate validation src/aegis_governance/mcp_server.py, src/engine/validation.py
Data protection / PII 12-field encryption src/telemetry/encryption.py
Cryptographic integrity Ed25519 + ML-DSA-44 hybrid src/crypto/hybrid_provider.py
Human-in-the-loop Two-key override, governance invariants src/workflows/override.py, CLAUDE.md 11.2
Audit trail completeness 100% logging, hash chain src/telemetry/emitter.py, src/telemetry/pipeline.py
Prompt injection defense Frozen params, schema validation schema/interface-contract.yaml

Summary

Both the CoSAI MCP Security Taxonomy and Red Hat's enterprise MCP architecture validate AEGIS's design approach. AEGIS already implements STRONG controls for 8 of 12 CoSAI threat categories, with MODERATE coverage on 3 and PARTIAL on 1 (transport security in HTTP mode). The primary gaps are operational (MCP request logging, rate limiting) rather than architectural.

The most significant validation: Red Hat's thesis that governance is the enterprise adoption bottleneck -- not protocol support -- directly positions AEGIS as the missing piece in the MCP ecosystem.


Sources

CoSAI

Red Hat

Supplementary