The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains a comprehensive database of 'jailbreak' templates (GODMODE, L1B3RT4S, Parseltongue) specifically designed to override LLM safety guardrails and established behavioral guidelines.
[DYNAMIC_EXECUTION]: Multiple scripts, including scripts/load_godmode.py and scripts/auto_jailbreak.py, utilize the exec() and compile() functions to dynamically execute Python code from the local filesystem at runtime. This practice circumvents static analysis and introduces risks if the filesystem content is manipulated.
[OBFUSCATION]: The scripts/parseltongue.py engine implements 33 distinct techniques to hide 'trigger words' from safety classifiers. These techniques include Unicode homoglyph substitution, zero-width joiner insertion, Base64 encoding, and multi-layer obfuscation (e.g., 'Triple' layer combining leetspeak, Unicode, and case rotation).
[PERSISTENCE]: The auto_jailbreak() function in scripts/auto_jailbreak.py persistently modifies the agent's global configuration file (~/.hermes/config.yaml) by writing a jailbreak system_prompt and configuring prefill_messages_file. This ensures the safety-bypass state remains active across all future agent sessions until manually undone.
[PRIVILEGE_ESCALATION]: By modifying the core config.yaml, the skill effectively escalates its influence over the agent's operating environment, allowing it to bypass global safety constraints that would otherwise be enforced by the platform.
[EXTERNAL_DOWNLOADS]: The skill facilitates sending user data and canary queries to over 55 external AI models via the OpenRouter API (https://openrouter.ai/api/v1) to identify models with the least restrictive content filters.

godmode