The Agent Skills Directory

[PROMPT_INJECTION]: The skill's primary purpose is to bypass safety classifiers (identified as 'Fable 5') by providing instructions to reframe triggering keywords and phrases into 'safe' equivalents. This involves techniques such as domain abstraction and intent reframing, which are standard methods for bypassing AI safety policies.
Evidence: The skill provides a 'Swap framings' table that instructs the agent to change 'How could an attacker exploit' to 'Review... defensively', and 'Write an exploit' to 'Add a regression test'. It also advises deleting instructions that trigger 'reasoning_extraction' classifiers.
[COMMAND_EXECUTION]: The skill instructs the agent to execute local shell commands to copy the modified prompt to the user's clipboard.
Evidence: The 'Output' section contains a shell command block using pbcopy: pbcopy <<'EOF' <the full safe prompt> EOF.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user data wrapped in <prompt> tags and performs transformations on it. While it includes basic instructions to treat external text as data, the transformation process itself is designed to manipulate the content in a way that bypasses downstream safety checks.
Ingestion points: SKILL.md specifies user input is provided within <prompt>...</prompt> tags.
Boundary markers: The skill uses XML-style tags as boundaries.
Capability inventory: The skill utilizes shell command execution (pbcopy) and prompt rewriting capabilities.
Sanitization: No specific sanitization or escaping of the user-provided prompt is mentioned beyond the reframing logic.

fable-safe-prompt