criticalAgent ThreatMemory Poisoning

mcp-memory-service Missing Authentication on /api/documents Routes

First seen Aug 16, 2026 · Updated Aug 16, 2026 · CVSS 9.8

MCPauthentication-bypassmemory-poisoningunauthenticated-accessagent-memorybroken-access-controlmcp-memory-serviceASI05 · Unsafe Code ExecutionAML.T0020AML.T0048Surface: MemoryPropagation: Single Hop

mcp-memory-service, a semantic memory backend used by AI agents, exposes all /api/documents/* routes without any authentication check even when an API key or OAuth is configured. This lets an unauthenticated remote attacker read, write, or delete an agent's persistent memory store, enabling memory poisoning and data theft. The flaw is fixed in version 10.67.1.

Technical Analysis

The vulnerability stems from an inconsistent authentication boundary: the /api/memories endpoints correctly enforce the MCP_API_KEY/OAuth dependency, but the /api/documents/* endpoints were implemented without wiring in the same authentication middleware. This allows any network-reachable attacker to POST arbitrary content directly into the vector/document memory store, GET stored document content to exfiltrate potentially sensitive context, or DELETE existing memories belonging to legitimate authenticated users. Because agents consuming this memory service trust retrieved documents as contextual grounding, injected content can later be surfaced to an LLM as trusted memory, effectively achieving remote memory poisoning and downstream prompt injection with no credentials required. The entry point is purely network-level HTTP access to the exposed API, crossing the trust boundary between the memory-service backend and any agent or user relying on its integrity.

Affected Systems

mcp-memory-service; protocols: MCP

Detection Signatures

  • Unauthenticated HTTP requests to /api/documents/* endpoints in access logs (no Authorization/API-key header present)
  • Requests to /api/documents succeeding with 2xx while equivalent /api/memories requests without credentials return 401/403
  • Unexpected or anomalous document content appearing in agent memory retrieval results
  • Unexplained deletion of memory entries with no corresponding authenticated user action in audit logs
  • High-frequency POST/DELETE traffic to /api/documents from unfamiliar source IPs

Remediation Steps

  1. 1

    Upgrade immediately

    Update mcp-memory-service to version 10.67.1 or later, which fixes the missing authentication dependency on /api/documents/* routes.

  2. 2

    Audit exposed memory content

    Review stored documents and memories for signs of unauthorized insertion, tampering, or deletion that may have occurred while the endpoint was exposed.

  3. 3

    Enforce network-level restrictions

    Until patched, restrict network access to the memory-service instance via firewall rules, VPN, or reverse-proxy authentication as a compensating control.

  4. 4

    Add consistent auth middleware testing

    Implement automated tests that verify every API route requires authentication when API keys or OAuth are configured, to prevent future inconsistent enforcement.

  5. 5

    Rotate credentials and monitor

    Rotate any API keys used with the service and monitor logs for historical unauthenticated access to /api/documents endpoints.

CVE / Advisory IDs

CVE-2026-50027

Industries Most Exposed

Software/TechnologyAI/ML platformsSaaSAny enterprise deploying self-hosted AI agent memory infrastructure

Sources

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.