mediumAgent ThreatPrivilege Abuse

Read-Only Bypass via Aggregation Pipeline in DocumentDB MCP Server

First seen Aug 6, 2026 · Updated Aug 6, 2026 · CVSS 5.5

MCPDocumentDBauthorization-bypassread-only-bypassaggregation-pipelineAWStool-misuseASI05 · Unsafe Code ExecutionSurface: Tool LayerPropagation: Single Hop

The AWS Labs DocumentDB MCP Server before version 1.0.12 fails to properly enforce read-only mode when processing certain aggregation pipeline stages, allowing an authenticated MCP client to perform unintended write operations on the connected database. This is a logic flaw in authorization enforcement rather than a novel AI-specific attack, but it is significant because agents and LLM-driven clients often rely on the read-only flag as a safety boundary to prevent destructive actions. Exploitation requires an already-authenticated client, limiting severity, but the impact on data integrity could be substantial in agentic workflows that assume read-only guarantees.

Technical Analysis

The DocumentDB MCP Server exposes an aggregation pipeline tool intended to allow read-only querying when configured in restricted mode. Certain aggregation stages (e.g., write-capable stages such as $merge or $out) are not properly filtered or validated by the server's read-only enforcement logic, letting an authenticated MCP client construct a pipeline that performs writes despite the read-only flag being set. This crosses the trust boundary between the MCP client (potentially an LLM agent operating under the assumption that read-only mode prevents state changes) and the underlying database, granting write/modify capability where only read access was intended. The attacker gains unauthorized data modification capability, which in agentic pipelines could lead to data corruption, unauthorized state changes, or serve as a stepping stone for further compromise if the agent's outputs or decisions depend on database integrity. No sandbox escape or cross-agent propagation is involved; impact is confined to the connected database resource.

Affected Systems

AWS Labs DocumentDB MCP Server; protocols: MCP

Detection Signatures

  • Aggregation pipeline requests containing write-capable stages ($merge, $out, $out-equivalent) submitted to an MCP server configured in read-only mode
  • Audit logs showing write operations (inserts/updates/collection creation) originating from sessions or API keys tagged as read-only
  • Monitor DocumentDB/MCP server logs for pipeline stage names inconsistent with declared read-only tool permissions
  • Alert on MCP tool invocations where declared tool capability metadata says 'read-only' but resulting database operation type is write

Remediation Steps

  1. 1

    Upgrade DocumentDB MCP Server

    Update to AWS Labs DocumentDB MCP Server version 1.0.12 or later, which fixes the authorization bypass in the aggregation pipeline tool.

  2. 2

    Enforce database-level access controls

    Do not rely solely on application-layer read-only flags; configure database user roles/permissions to be genuinely read-only at the DocumentDB IAM/role level as defense in depth.

  3. 3

    Audit and validate pipeline stages

    Implement server-side allowlisting of permitted aggregation stages for read-only sessions, explicitly rejecting write-capable stages like $merge and $out.

  4. 4

    Review logs for exploitation

    Audit historical MCP server logs for any write operations executed under sessions configured as read-only to identify potential prior exploitation.

CVE / Advisory IDs

CVE-2026-18954

Industries Most Exposed

Cloud/SaaSTechnologyFinancial ServicesHealthcareAny organization using AWS DocumentDB with MCP-based AI agent tooling

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.