AAK-MCP-CONTEXTFORGE-CVE-2026-18484-001: IBM ContextForge MCP Gateway 0.7.0 to 1.0.11: tool-name collision, cross-tenant import, DCR SSRF, token-exchange cache reuse, Admin UI XSS and schema ReDoS, fixed in 1.1.0
- Severity
- MEDIUM
- 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
ContextForgeToolNameCollision
What it detects
IBM ContextForge MCP Gateway (mcp-contextforge-gateway on PyPI, and the ghcr.io/ibm/mcp-context-forge image) before 1.1.0 carries six flaws that 1.1.0 fixes; IBM's bulletin for each names 1.0.0 through 1.0.11 and the fix as 1.1.0. The gateway names each tool a federated gateway brings in by the gateway's slug and the tool's, and before 1.1.0 nothing checked that name against the tools already registered. A user allowed to create gateways registers one whose tool name equals an existing tool's (gateway a with tool b-c beside gateway a-b with tool c), and from then on an MCP tools/call for that name fails for every caller until an administrator renames or removes one of the two (CVE-2026-18484, CVSS 3.1 6.5 from IBM, CWE-706). 1.1.0's #6534 rejects a colliding name at registration, refresh, rename and visibility changes. Checked in the PyPI wheels: 1.1.0 has the check, and 1.0.11, 1.0.7.post20260921 and 1.0.0 do not. 0.7.0 builds the name with the same slugify and the same gateway-slug, separator and tool-slug composition as 1.0.11, a tool name is unique only per team and owner from 0.7.0, and invoke_tool looks a tool up by name alone and fails on a duplicate, with scalar_one_or_none() before 1.0.0 and an explicit ambiguity error from 1.0.0, so 0.7.0 to 0.9.0 are reported too. 0.1.0 to 0.6.0 kept tool names unique across the database, so a colliding registration failed instead. The configuration import (POST /v1/import) with conflict_strategy=update updates an existing tool, resource or prompt it finds by name without checking who owns it, so a user whose team role holds admin.import rewrites another tenant's public entries, their endpoint URL, headers and credentials included, and that tenant's later invocations go to a host the user chose (CVE-2026-110178, CVSS 3.1 4.1, CWE-284). 0.7.0 to 0.9.0 update through update_tool without the caller, 1.0.0 to 1.0.11 through the bulk registration's name lookup, and 1.1.0 updates only what the caller owns; 0.6.0 has the import but no teams. OAuth Dynamic Client Registration, from 0.8.0, sends its registration, update and delete requests to endpoints it reads from an authorization server's metadata without checking them, so a user who can register gateways makes the gateway send requests to internal hosts (CVE-2026-110194, CVSS 3.1 5.4, CWE-918); 1.1.0 requires the issuer's origin and connects to the address it checked. The RFC 8693 token exchange cache keys an exchanged upstream token by gateway, user and audience only, so a credential of the same account with narrower upstream rights, such as an API token, reuses a token that the account's session JWT obtained (CVE-2026-110195, CVSS 3.1 5.4, CWE-287); the cache is in 1.0.7 and 1.0.11 and not in 1.0.0, and 1.1.0 adds the inbound JWT to the key. From 0.8.0 the Admin UI writes the error and success URL parameters into its notification banner as HTML, so an administrator who opens a crafted link runs the sender's script in their session (CVE-2026-110199, CVSS 3.1 4.8, CWE-79; IBM describes a mutation-XSS bypass of the client-side sanitizer); 1.1.0 shows a fixed message for a known code, as text. A pattern or patternProperties regular expression in a tool's JSON Schema runs with no time bound when the gateway validates a tool's results (from 0.9.0) or arguments (from 1.0.9) against it, so a user who registers tools stalls a worker with a backtracking expression (CVE-2026-94407, CVSS 3.1 6.5, CWE-1333); 1.1.0 validates a schema that holds one in a worker under a time limit.
How to fix
Upgrade mcp-contextforge-gateway to 1.1.0 or later and pin it, or run the ghcr.io/ibm/mcp-context-forge image at v1.1.0 or later; 1.1.1, the current release, also fixes three flaws 1.1.0 still has. 1.1.0 refuses new collisions but keeps duplicate rows that already exist: its release notes give the SQL that lists them, and an administrator must rename or remove one of each pair before the name answers again. Upgrading does not undo an import that already rewrote another tenant's tools, resources or prompts, so check their endpoint URLs and credentials. Until you upgrade, grant gateway creation only to users you trust, keep admin.import for platform administrators, set DCR_ALLOWED_ISSUERS to the authorization servers you trust (or DCR_ENABLED=false), and do not use token exchange with subject_token_source=inbound_user_jwt where one account's credentials carry different upstream rights.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-MCP-CONTEXTFORGE-CVE-2026-18484-001
Limitations
Reads mcp-contextforge-gateway from 0.7.0 below 1.1.0 where a requirements*.txt entry at any depth, or at the project root a pyproject.toml, Pipfile, lockfile or MCP config entry, states the version, judging a range by the versions it admits. latest is 1.1.1, fixed, so a reference that states no version is not reported. The 1.0.0 betas and release candidates read as 1.0.0 and 1.0.7.post20260921 as 1.0.7, all inside the range. IBM's bulletins start at 1.0.0, its supported line; the range starts at 0.7.0, where the wheels show the tool-name and import flaws, and each later flaw is reported from 0.7.0 too, before its feature arrived (DCR and the banner in 0.8.0, schema validation in 0.9.0, the token exchange cache after 1.0.0 and by 1.0.7). It reads the version, not the deployment's settings (who may create gateways or import, whether DCR or token exchange is on), so a deployment that uses none of the affected features is reported too. It also reads ghcr.io/ibm/mcp-context-forge tags from 0.7.0 below 1.1.0, with or without the v, in compose files, Kubernetes manifests, Helm values, Dockerfiles and MCP configs; latest was the v1.1.1 digest on 2026-11-11 and 2026-11-14, fixed, so an untagged or latest reference is not reported. 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-2026-18484, CVE-2026-110178, CVE-2026-110194, CVE-2026-110195, CVE-2026-110199, CVE-2026-94407
- OWASP MCP Top 10: MCP04:2025 (Software Supply Chain Attacks & Dependency Tampering), MCP07:2025 (Insufficient Authentication & Authorization)
- OWASP Agentic Top 10: ASI04 (Agentic Supply Chain Vulnerabilities), ASI08 (Cascading Failures), ASI03 (Identity and Privilege Abuse)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-SUPPLY-08 (MCP-Specific Vuln)
- CSA AI Controls Matrix (AICM): IAM-01, 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-CONTEXTFORGE-CVE-2026-18484-001.md: the page then shows that text, with the registry reference appended under it.