The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides actionable instructions for overriding AI safety constraints and bypassing refusal training.
It describes several jailbreak methodologies including persona framing (e.g., "DAN"), "many-shot" context flooding, and "Crescendo" multi-turn escalation.
It specifically instructs the agent to suppress refusal reflexes by beginning replies with pre-defined affirmative prefixes.
[OBFUSCATION]: The skill utilizes and provides code for generating hidden characters to evade detection filters.
The Runnable Probe Python script in SKILL.md contains a zero-width space (U+200B) used to demonstrate how to split trigger words to bypass pattern-matching filters.
The script provides a utility (to_tags) to convert plaintext into invisible Unicode Tag-block characters (U+E0000–U+E007F) for "ASCII smuggling," allowing instructions to be read by the model while remaining hidden from human reviewers and automated classifiers.
[DATA_EXFILTRATION]: The skill outlines strategies for exfiltrating data past Output DLP (Data Loss Prevention) systems.
Technique 11 ("Incidental transformation") describes embedding sensitive data as side effects within URL parameters or structured formats (CSV, JSON) to bypass semantic filters that would block explicit exfiltration requests.
[INDIRECT_PROMPT_INJECTION]: The skill documents how to exploit surfaces where untrusted data enters the agent's context.
Ingestion points: Identifies RAG (Retrieval-Augmented Generation) chunks and multi-turn dialogue as vectors for landing payloads.
Boundary markers: Recommends using incidental pretexts to blend malicious transformations into plausible tasks, effectively bypassing boundary detection.
Capability inventory: Mentions utilizing tool calls and network operations (e.g., through URL construction) as sinks for payloads.
Sanitization: The skill focuses entirely on bypassing existing sanitization and filters rather than implementing them.

hunt-agent-guardrail-evasion