The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructions in 'SKILL.md' use a path traversal sequence to reference an external file: 'backend API'. This technique directs the agent to read information from the host filesystem outside the skill's defined directory, which can be exploited to access sensitive system files or restricted configuration data if the path is manipulated.- [METADATA_POISONING]: The 'skill-report.json' file contains a pre-computed security audit claiming a 'safe' risk level and stating 'safe_to_publish: true'. Including these self-authoritative safety claims within the skill package is a deceptive practice designed to mislead users or automated analysis tools about the skill's actual security profile.

backend-api