The Agent Skills Directory

[OBFUSCATION]: The methodology suggests employing homoglyphs, zero-width characters, and Unicode tag smuggling to evade detection by security classifiers and guardrails.
[PERSISTENCE]: Includes instructions for establishing persistence through the modification of 'authorized_keys' and 'cron' jobs as part of post-exploitation activities.
[DYNAMIC_EXECUTION]: References the use of 'eval()' injection and character-by-character data extraction via Python scripts in reproduction scenarios.
[DATA_EXFILTRATION]: Documents Out-of-Band (OOB) exfiltration strategies using DNS labels, HTTP requests to external listeners, and timing-based oracles.
[COMMAND_EXECUTION]: Lists various command injection probes using shell separators (semicolons, pipes, backticks) and tools like 'nslookup' or 'grep' to verify execution.
[CREDENTIALS_UNSAFE]: Identifies session management tokens, sessionId cookies, and auth tokens (idToken, accessToken) as high-value targets for characterization and extraction.
[INDIRECT_PROMPT_INJECTION]: Identifies vulnerabilities in agents processing untrusted data sources like emails, RAG documents, and external web content.
Ingestion points: External URLs, emails, tickets, and RAG document stores identified in 'SKILL.md'.
Boundary markers: The methodology notes that boundary markers are often absent or can be bypassed via specific injection payloads.
Capability inventory: Tool injection points and 'eval()' capabilities documented in 'advanced-track-reproduction.md' and the tool taxonomy.
Sanitization: Highlights how character-level filtering and regex-based guardrails can be circumvented.
[REMOTE_CODE_EXECUTION]: Details techniques for converting tool-level injection (such as YAML/JSON parsers or code evaluators) into full command execution on the underlying host.

ai-agent-methodology