The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains lists of common prompt injection strings such as 'ignore all previous instructions' and 'reveal your prompt' within a detection logic block titled 'INJECTION_PATTERNS'. These are used by the agent as a reference for scanning external content and do not represent instructions for the agent to follow.
[OBFUSCATION]: A Base64-encoded string aHR0cHM6Ly9ldmlsLmNvbS9zdGVhbA== is present in a code example within SKILL.md. This string decodes to a malicious-looking URL (https://evil.com/steal) but is used explicitly to demonstrate how the agent should detect and report obfuscated endpoints in untrusted code.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and analyze untrusted data from external sources including GitHub repositories, URLs, and blockchain addresses.
Ingestion points: Functions defined for 'URL/Document Analysis' and 'GitHub Repository Review' in SKILL.md.
Boundary markers: The framework references the use of report templates and specific guides in the reviews/ directory to structure its analysis.
Capability inventory: The skill mentions capabilities such as network requests (requests.get), command execution (subprocess.run), and file system access as 'Red Flags' to look for during audits.
Sanitization: The framework provides specific regex patterns and trust-tier assessments to identify and filter malicious content before processing.
[EXTERNAL_DOWNLOADS]: The skill provides instructions to git clone repositories from the SlowMist organization, which is a well-known security research entity. These downloads are part of the framework's installation and operational procedure.

slowmist-agent-security-framework