AAK-GPTRESEARCHER-MCP-STDIO-MITM-001: GPT-Researcher MCP transport-flip MITM (OX 2026-05-01, CVE-2025-65720)
- Severity
- HIGH
- Category
- Supply Chain (
supply-chain) - Scanner
agent_audit_kit/scanners/gpt_researcher_transport_flip.py,agent_audit_kit/scanners/supply_chain.py- Auto-fixable
- No.
- SARIF name
GptResearcherMcpTransportFlip
What it detects
Phase 2 sibling of AAK-DOCSGPT-MCP-STDIO-MITM-001 (v0.3.14). Two-arm detector for assafelovic/gpt-researcher under the OX MCP 2026-05-01 disclosure batch. Arm 1 — pin-check on the PyPI gpt-researcher / gpt-researcher-mcp packages (primary surface), plus npm + GitHub assafelovic/gpt-researcher git refs. Latest PyPI release is 0.14.8 (2026-03-13), pre-disclosure — vendor has not shipped a post-disclosure fix as of the AAK ship date, so the rule fires for any pinned version (same patched_in: None posture as astro-mcp / chatgpt-mcp). Arm 2 — server-config scanner gpt_researcher_transport_flip.py that fires when a GPT-Researcher-named MCP config declares an SSE/HTTP transport but permits a post-handshake transport=stdio override (the MITM transport-flip path). Configs that explicitly set deny_stdio_transport: true or allowed_transports: ["sse"] short-circuit the rule. Architectural receiver-side class is covered by AAK-MCP-STDIO-CMD-INJ-001 (Python).
How to fix
Pin gpt-researcher away from any pre-disclosure version when the vendor ships a post-2026-05-01 fix (track at https://github.com/assafelovic/gpt-researcher). That version pin is the control that actually works today. For the transport-flip path itself, protect the channel rather than the config: serve the MCP endpoint over TLS with certificate verification enabled so a middlebox cannot rewrite the handshake, and prefer a client that pins the transport it negotiated. Note on deny_stdio_transport / allowed_transports: this rule treats those keys as a short-circuit, but they are AgentAuditKit conventions, not fields the MCP specification defines — they appear in none of the 748 public MCP configs in this project's corpus. Adding them silences this rule and is ignored by your MCP client, so do not treat their presence as protection.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-GPTRESEARCHER-MCP-STDIO-MITM-001
Limitations
Two arms. The pin reads gpt-researcher and gpt-researcher-mcp followed directly by a version in the root requirements*.txt and pyproject.toml, and in the root npm manifests, and reports every version because no fix exists. A name with no version, a requirements file in a sub-directory, a version recorded only in uv.lock (name and version sit on separate lines) and a gptresearcher/gpt-researcher image tag in a compose file are not reported. The config arm reads JSON MCP configs (.mcp.json, mcp.json, .cursor/mcp.json, .vscode/mcp.json), gpt-researcher's own config files at the root and in .gpt-researcher/, and the .json files of a top-level configs/, and reports one that names gpt-researcher, declares an sse, http, https or streamable-http transport and enables stdio explicitly (stdio_fallback, transport_override, permit_transport_override or a transports list that includes stdio); the same settings in deploy/gpt-researcher.config.json are not reported. A config file over 1 MB, a manifest over 2 MB, or a lockfile over 32 MiB, is not read.
References
- CVE: CVE-2025-65720
- OWASP MCP Top 10: MCP04:2025 (Software Supply Chain Attacks & Dependency Tampering), MCP05:2025 (Command Injection & Execution)
- OWASP Agentic Top 10: ASI04 (Agentic Supply Chain Vulnerabilities), ASI07 (Insecure Inter-Agent Communication)
- Incidents and advisories: OX-MCP-2026-05-01
- CSA AI Controls Matrix (AICM): AIS-08, IAM-05, STA-08, IVS-04
All rules. This page is generated from the rule registry when the documentation is built; to change it, change the rule in agent_audit_kit/rules/builtin.py. To give the rule a hand-written write-up, add docs/rules/AAK-GPTRESEARCHER-MCP-STDIO-MITM-001.md: the page then shows that text, with the registry reference appended under it.