Vulnerability  ·  2026-09-04

Tencent AI-Infra-Guard skill-scan excludes compiled Python bytecode from malicious-skill scanning

VulnerabilityMedium impactGlobalCVE-2026-84809
Published Sept 2, 2026 (CVSS 6.5), this is a scanner-evasion vulnerability in a security tool specifically designed to vet AI agent 'skills' for malicious content, undermining its core function.
As the AI-agent-skill ecosystem grows (analogous to npm/PyPI packages but for agent capabilities), security scanners like AI-Infra-Guard are a key supply-chain defense; a trivial bypass (hiding payloads in .pyc files) means organizations trusting this scanner's clean bill of health could unknowingly deploy malicious agent skills.
The skill-scan component hardcodes __pycache__ directories and .pyc/.pyo/.pyd extensions into skip lists across multiple scanning surfaces. An attacker can distribute an agent 'skill' bundle with benign-looking Python source files alongside a malicious compiled bytecode file; the scanner never inspects the bytecode, allowing malicious code to reach execution undetected.
Tencent AI-Infra-Guard skill-scan component
Await vendor patch; in the interim, do not rely solely on AI-Infra-Guard skill-scan for supply-chain vetting of agent skills — manually inspect or disallow compiled bytecode in skill packages.
NVDGitHub - Tencent/AI-Infra-Guard
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →