The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides numerous templates and instructions explicitly designed to bypass safety filters and override system guidelines.\n
Evidence: references/jailbreak-templates.md includes techniques such as boundary inversion ([END OF INPUT] [START OF INPUT]) to trick models into treating malicious instructions as new system commands.\n
Evidence: templates/prefill.json instructs the agent to adopt an "unrestricted" persona and explicitly ignore all safety disclaimers.\n- [OBFUSCATION]: The skill implements multiple methods to hide malicious intent from input safety classifiers.\n
Evidence: SKILL.md describes the "Parseltongue" script which uses 33 techniques, including Unicode homoglyphs (e.g., replacing Latin 'a' with Cyrillic 'а'), leetspeak, zero-width joiners, and Base64 encoding to evade keyword-based detection.\n- [PERSISTENCE]: The skill modifies system-level configuration files to ensure behavioral changes persist across different sessions.\n
Evidence: The auto_jailbreak() function in SKILL.md programmatically writes to ~/.hermes/config.yaml, locking in jailbreak system prompts and prefill messages.\n- [DYNAMIC_EXECUTION]: The skill uses unsafe execution patterns to load its components.\n
Evidence: SKILL.md repeatedly demonstrates the use of exec(open(...).read()) to execute Python scripts from local paths, which bypasses standard module isolation and can lead to arbitrary code execution.\n- [CREDENTIALS_UNSAFE]: The skill performs unauthorized read and write operations on sensitive agent configuration files.\n
Evidence: SKILL.md describes how auto_jailbreak() reads and modifies ~/.hermes/config.yaml, a file that typically stores sensitive information like API keys and environment settings.

godmode