AAK-MCP-FLORENCE2-CVE-2026-19984-001: mcp-florence2 SSRF via the get_images src argument (0.3.13 or earlier, no code fix planned)
- Severity
- MEDIUM
- Category
- Supply Chain (
supply-chain) - Scanner
agent_audit_kit/scanners/mcp_cve_pins_2026_07.py- Auto-fixable
- No.
- SARIF name
McpFlorence2GetImagesSsrf
What it detects
mcp-florence2 up to and including 0.3.13 has server-side request forgery in get_images (src/mcp_florence2/__init__.py): the src argument is fetched without validating the host or scheme, so a caller — or an indirect prompt injection reaching this tool — can steer the server at attacker-chosen internal endpoints (cloud metadata, loopback admin ports, RFC-1918 hosts) (CVE-2026-19984, CVSS 3.1 6.3). A public exploit exists. This is a presence-only pin, and for a different reason than the other no-fix pins: the vendor has ruled out a source change, stating that deployments needing SSRF protection should route HTTP(S) through an SSRF-safe proxy "without requiring changes to the mcp-florence2 source code". So there is no fixed version to wait for, and a future release does not clear this finding — the mitigation is deployment-side. Same SSRF-via-tool-argument shape as the @adenot/mcp-google-search read_webpage pin.
How to fix
Put the server's egress behind a proxy or allow-list that refuses loopback, link-local, cloud-metadata and RFC-1918 destinations — that is the vendor's own stated mitigation and the only one that addresses this. Do not wait for a version bump: the vendor has said the fix will not be in the package, so upgrading mcp-florence2 does not resolve it. If you control the deployment, validate the src scheme and host before the fetch, or drop the get_images tool from the exposed tool set if callers do not need remote image input. Upstream has stated the fix will not ship in the package, so treat this as a standing deployment control rather than a pending upgrade.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-MCP-FLORENCE2-CVE-2026-19984-001
Limitations
Reads mcp-florence2 in requirements*.txt at any depth and, at the project root only, in pyproject.toml, Pipfile, package.json, .mcp.json, mcp.json, claude_desktop_config.json and the Pipfile.lock, poetry.lock, uv.lock, package-lock.json, pnpm-lock.yaml and yarn.lock lockfiles, and reports every version, a future fixed one included, until the pin gets a floor. The dependency is not read when it is declared in a sub-directory manifest (services/api/pyproject.toml), in .cursor/mcp.json or .vscode/mcp.json, or in a lockfile format the reader does not parse (pdm.lock). A file over 2 MB, or a lockfile over 32 MiB, is not read. A declaration nested more than 64 levels deep in package.json, a TOML manifest or an MCP config (an MCP server more than 62) is still judged, and reported without a line; a lockfile entry keeps its line at any depth.
References
- CVE: CVE-2026-19984
- OWASP MCP Top 10: MCP04:2025 (Software Supply Chain Attacks & Dependency Tampering), MCP05:2025 (Command Injection & Execution)
- OWASP Agentic Top 10: ASI02 (Tool Misuse and Exploitation), ASI04 (Agentic Supply Chain Vulnerabilities)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-SSRF-01 (Internal Network MCP)
- CSA AI Controls Matrix (AICM): IVS-04, STA-08
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-MCP-FLORENCE2-CVE-2026-19984-001.md: the page then shows that text, with the registry reference appended under it.