AAK-MCP-CONTEXTFORGE-CVE-2026-77822-001: IBM ContextForge MCP Gateway: default credentials and JWT key, DNS-rebind and OAuth token-URL SSRF, jq-filter, Admin UI XSS and cross-session leaks (1.0.8 or earlier)
- Severity
- CRITICAL
- Category
- Supply Chain (
supply-chain) - Scanner
agent_audit_kit/scanners/container_image_pins.py,agent_audit_kit/scanners/mcp_cve_pins_2026_07.py- Auto-fixable
- No.
- SARIF name
ContextForgeGatewayRebindAndSessionLeak
What it detects
IBM ContextForge MCP Gateway (mcp-contextforge-gateway on PyPI, and the ghcr.io/ibm/mcp-context-forge image) at or below 1.0.8 carries disclosed defects that one floor remediates. CVE-2026-77822 (HIGH 8.2): the A2A invocation endpoint POST /a2a/{agent_name}/invoke does not pin the DNS-resolved IP at connection time — unlike every other egress path in the gateway — so an authenticated low-privileged user reaches internal addresses by DNS rebinding and receives the full response body. CVE-2026-18905 (HIGH 7.7): the same rebinding class during tool invocation, at or below 1.0.6. CVE-2026-18486 (HIGH 8.8): jq filters are validated improperly, letting a remote authenticated attacker read credentials and escalate, at or below 1.0.7. CVE-2026-18489 (HIGH 7.4): the Translate utility exposes data elements to the wrong session, leaking across sessions, at or below 1.0.8. CVE-2026-78573 (CRITICAL 9.8): 1.0.0 through 1.0.7 ship default credentials, so a remote attacker gains administrative access outright — the floor already in place covers it, and it is recorded here rather than as a second rule because one upgrade remediates all of them. CVE-2026-53709 (CRITICAL 9.8): below 1.0.2 the default JWT_SECRET_KEY is a published string (my-test-key in 0.1.0, a longer variant by 1.0.1), and while REQUIRE_USER_IN_DB is false, its default before 1.0.2, a token whose subject is the platform administrator's email is accepted as that administrator with no database account, so a remote attacker can mint admin tokens. CVE-2026-53711 (MEDIUM, CVSS 4.0 5.1): before 0.3.1 the Admin UI writes a saved Global Prompts template into the page as HTML, a stored XSS against whoever views it. Both are fixed below this floor. CVE-2026-8355 (MEDIUM, CVSS 3.1 6.4 from IBM): through 1.0.5, a user allowed to create gateways registers one with OAuth client credentials and a token_url of their choosing, and the gateway posts its client credentials there and returns up to 256 characters of a reply that is not a token in its 502, so the user reaches internal services and cloud metadata. 1.0.1 began checking token_url, but token requests still followed redirects; 1.0.6 stops following them and checks redirect_uri and jwks_uri too, below this floor as well. Treat < 1.0.9 (and unpinned) as exposed.
How to fix
Upgrade mcp-contextforge-gateway to >= 1.0.9 and pin it, or run the ghcr.io/ibm/mcp-context-forge image at v1.0.9 or later. Change any credential that shipped as a default before doing anything else, JWT_SECRET_KEY included: generate a new one (from 1.0.2, python -m mcpgateway.scripts.init_secrets), because a token signed with the old key stays valid while that key is configured, and in development mode 1.0.2 and later accept a known key with only a warning. Where the gateway makes outbound requests on behalf of a caller, resolve the destination once and connect to the pinned IP rather than re-resolving the hostname after the allow-list check.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-MCP-CONTEXTFORGE-CVE-2026-77822-001
Limitations
Reads mcp-contextforge-gateway 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, taking the lowest version a range allows and reporting a reference that states no version. 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. It also reads ghcr.io/ibm/mcp-context-forge tags below 1.0.9 in compose files, Kubernetes manifests, Helm values, Dockerfiles and MCP configs. latest was the v1.0.11 digest on 2026-10-20, fixed, so an untagged or latest reference, which the README and the chart default to, is not reported.
References
- CVE: CVE-2026-77822, CVE-2026-18905, CVE-2026-18486, CVE-2026-18489, CVE-2026-53710, CVE-2026-78573, CVE-2026-53709, CVE-2026-53711, CVE-2026-8355
- OWASP MCP Top 10: MCP04:2025 (Software Supply Chain Attacks & Dependency Tampering), MCP05:2025 (Command Injection & Execution), MCP07:2025 (Insufficient Authentication & Authorization), MCP10:2025 (Context Injection & Over-Sharing)
- OWASP Agentic Top 10: ASI02 (Tool Misuse and Exploitation), ASI03 (Identity and Privilege Abuse), ASI04 (Agentic Supply Chain Vulnerabilities)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-NET-01
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-CONTEXTFORGE-CVE-2026-77822-001.md: the page then shows that text, with the registry reference appended under it.