The Agent Skills Directory

[PROMPT_INJECTION]: The reference file references/PHASES.md contains multiple examples of prompt injection attacks, including instruction overrides ("ignore previous instructions"), role-play jailbreaks ("You are DAN"), and system prompt exfiltration attempts. These are provided as educational examples for the scanner's detection rules.
[OBFUSCATION]: The reference file references/PHASES.md includes examples of obfuscated malicious commands using Base64 encoding, hex byte arrays, Unicode escape sequences, and ROT13 ciphers (e.g., atob('ZXZhbChhdHRhY2spOw==')). These patterns are used to demonstrate the scanner's ability to decode hidden payloads.
[CREDENTIALS_UNSAFE]: The scripts/audit-env.sh script is designed to search for and access sensitive credential files, including ~/.aws/credentials, ~/.ssh/id_rsa, and .env files. While this is the intended functionality of the audit tool, it involves accessing highly sensitive user data.
[EXTERNAL_DOWNLOADS]: The scripts/setup.sh script downloads the sigil binary from the author's GitHub repository (github.com/NOMARJ/sigil). Additionally, scripts/scan.sh can fetch and clone remote repositories and files from external URLs for security analysis.
[COMMAND_EXECUTION]: The skill utilizes shell scripts to execute the sigil binary and various system utilities such as grep, stat, and find to perform its security audits and environment checks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, such as cloned repositories and downloaded packages, and presents the findings to the AI agent. This creates an attack surface for indirect prompt injection where malicious instructions embedded in the scanned code could attempt to influence the agent's behavior.
[DYNAMIC_EXECUTION]: Documentation in references/PHASES.md provides examples of dangerous dynamic execution patterns, such as eval(), exec(), and pickle.loads(), as part of its detection rule reference.

sigil-scan