AAK-MCP-OPENCLAW-CVE-2026-62195-001: OpenClaw MCP authorization and network-policy bypasses (below 2026.7.1)
- Severity
- HIGH
- Category
- Supply Chain (
supply-chain) - Scanner
agent_audit_kit/scanners/mcp_cve_pins_2026_07.py- Auto-fixable
- No.
- SARIF name
OpenClawMcpLoopbackAuthzBypass
What it detects
OpenClaw 2026.5.20 up to (but not including) 2026.6.6 contains an authorization-bypass in the MCP loopback feature that lets lower-trust callers execute owner-only tools by routing through configured input paths, so an attacker can execute or persist actions beyond their intended permissions (CVE-2026-62195, CVSS 8.3; NVD range 2026.5.20 <= v < 2026.6.6). The same pin also covers CVE-2026-62208 (Authorization headers forwarded during MCP SSE redirects, fixed in 2026.6.5). Before 2026.7.1, owner-only authorization is also not enforced for Claude Code permission prompts delivered through the MCP channel bridge: a non-owner channel sender with channel command access can approve or deny a pending permission request meant for the owner, so the requested action proceeds without owner consent (CVE-2026-100585, CVSS 8.0, CWE-862; GHSA-p5g8-m35v-7m82 scopes it < 2026.7.1 with no lower bound). Before 2026.7.1 a non-owner channel sender can also run the Codex computer-use install command, and so install plugins and start MCP processes with OpenClaw's privileges (CVE-2026-100587, CVSS 8.8, CWE-862, GHSA-pjjr-5qhr-5w6r). The same sender can persist a stdio MCP command through /mcp set or /mcp unset that runs with OpenClaw's privileges when the configuration loads (CVE-2026-100596, CVSS 8.8, CWE-862, GHSA-wwx7-573h-pqwc). Before 2026.6.6 the browser tool's CDP discovery also accepted WebSocket URLs that OpenClaw's network policy blocks, so a lower-trust caller or configured input path could reach destinations the policy should have blocked (CVE-2026-62197, CVSS 8.5, CWE-918, GHSA-3x84-qq85-fj65). Treat every release below 2026.7.1 (and unpinned) as exposed.
How to fix
Upgrade openclaw to >= 2026.7.1 and pin it. Enforce the caller's trust tier on every MCP loopback tool invocation, not just at the transport edge, and accept a permission reply only from the owner the prompt was addressed to.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-MCP-OPENCLAW-CVE-2026-62195-001
Limitations
Reads the npm package openclaw at the project root, in package.json, .mcp.json, mcp.json, claude_desktop_config.json and the 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 PyPI project openclaw, an unrelated installer for the cmdop CLI, is a different package and is not read. The dependency is not read when it is declared in a sub-directory manifest (packages/app/package.json) or in a lockfile format the reader does not parse (a Yarn 2+ yarn.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-2026-62195, CVE-2026-62208, CVE-2026-100585, CVE-2026-100587, CVE-2026-100596, CVE-2026-62197
- OWASP MCP Top 10: MCP04:2025 (Software Supply Chain Attacks & Dependency Tampering), MCP07:2025 (Insufficient Authentication & Authorization)
- OWASP Agentic Top 10: ASI03 (Identity and Privilege Abuse), ASI04 (Agentic Supply Chain Vulnerabilities)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-AUTH-01 (Missing Authentication)
- 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-OPENCLAW-CVE-2026-62195-001.md: the page then shows that text, with the registry reference appended under it.