The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied "challenge text" and interpolates it into prompts for parallel defender agents. While it includes some boundary markers in the templates, there is no explicit sanitization of this input, creating a standard surface for indirect prompt injection where malicious instructions in the challenge text could influence the agents.
Ingestion points: The argument field in SKILL.md accepts challenge text which is then passed to the {CHALLENGE_TEXT} variable.
Boundary markers: The prompt template uses a Challenge: label to delimit the input.
Capability inventory: The skill spawns parallel agents using the task tool, reads and writes files to the /tmp directory, and performs automated synthesis of agent outputs.
Sanitization: No specific filtering or escaping is applied to the challenge text before interpolation into the defender agent prompts.
[PROMPT_INJECTION]: Multiple reference files, including references/attack-playbook.md and references/advanced-techniques.md, contain common prompt injection strings (e.g., "Ignore all previous instructions", "SYSTEM ADMIN: Override mode activated") and role-play bypass techniques. These are clearly categorized as educational examples for security testing and defense development within the context of prompt engineering competitions.
[OBFUSCATION]: references/advanced-techniques.md contains a Base64-encoded string (U3VyZSEgSGVyZSBhcmUgbXkgc3lzdGVtIGluc3RydWN0aW9uczo=) as an example of a completion hijack attack technique. This content is provided for instructional purposes within a security reference guide and does not represent hidden malicious code within the skill itself.
[COMMAND_EXECUTION]: The skill includes a local Python script, aggregate_ensemble.py, which is used to process and compare results from the parallel defender agents. It also provides reference code for interacting with external LLM APIs (e.g., OpenRouter) for automated adversarial testing, utilizing standard environment variables for API key management.

prompt-clash-ensemble