AAK-MCP-TOOLUNIVERSE-CVE-2026-81096-001: ToolUniverse unauthenticated RCE via python_code_executor sandbox escape (1.2.6 or earlier)
- Severity
- CRITICAL
- Category
- Supply Chain (
supply-chain) - Scanner
agent_audit_kit/scanners/mcp_cve_pins_2026_07.py- Auto-fixable
- No.
- SARIF name
ToolUniversePythonExecutorSandboxEscape
What it detects
ToolUniverse (tooluniverse, PyPI) through 1.2.6 runs caller-supplied Python inside a sandbox that can be escaped, on a server that requires no authentication (CVE-2026-81096, CVSS 10.0). The executor behind the python_code_executor tool inspects submitted source against a deny-list of attribute names and calls but leaves the attribute-lookup builtins available, so a dunder attribute reached through a string lookup — or through a module already permitted — lets a caller walk from a literal's class to its base and enumerate subclasses until it holds a reference to subprocess. A per-call argument also widens the import allow-list before the inspection runs. The HTTP and MCP servers (http_api_server.py, smcp_server.py) bind every interface with debugging enabled and no authentication, so any caller that reaches the port executes code as the server process. 1.3.0 adds bearer-token authentication, defaults the bind address to loopback, and hardens the attribute checks. Treat < 1.3.0 (and unpinned) as exposed.
How to fix
Upgrade tooluniverse to >= 1.3.0 and pin it. Do not expose the code-executor tool on a non-loopback interface, and require authentication in front of it — a deny-list over attribute names is not a sandbox boundary.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-MCP-TOOLUNIVERSE-CVE-2026-81096-001
Limitations
Dependency-level detection only: this fires on the pinned or unpinned tooluniverse reference, not on the sandbox-escape pattern itself. The pattern itself is covered by AAK-SANDBOX-DENYLIST-001, which shipped for issue #704 and fires on a deny-list of names checked against a value that is then executed in-process with a lookup primitive still reachable. 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-2026-81096
- OWASP MCP Top 10: MCP04:2025 (Software Supply Chain Attacks & Dependency Tampering), MCP05:2025 (Command Injection & Execution), MCP07:2025 (Insufficient Authentication & Authorization)
- OWASP Agentic Top 10: ASI04 (Agentic Supply Chain Vulnerabilities), ASI05 (Unexpected Code Execution (RCE))
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-EXEC-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-TOOLUNIVERSE-CVE-2026-81096-001.md: the page then shows that text, with the registry reference appended under it.