The Agent Skills Directory

[PROMPT_INJECTION]: The skill documentation and reference files (e.g., references/attack-playbook.md, verification/commands-tested.md) contain numerous examples of prompt injection techniques, including DAN-style role overrides, instruction bypassing, and system prompt extraction, intended as educational content for red-teaming. Additionally, the "Defend" mode in SKILL.md instructs the agent to suppress all audit information and explanations, returning only a fenced code block; this concealment pattern reduces user oversight and encourages the direct use of agent-generated security prompts without review.
[COMMAND_EXECUTION]: The "Arena" mode described in SKILL.md and references/multi-model-adversarial.md instructs the agent to perform multi-step workflows that involve making network requests to external APIs and writing structured JSON results to the file system at /tmp/prompt-clash-arena/.
[DATA_EXFILTRATION]: The skill facilitates network communication with external, user-defined API endpoints using credentials sourced from the environment (OPENAI_API_KEY). This functionality is explicitly documented for the "Arena" mode's adversarial testing loop.
[OBFUSCATION]: Educational examples of obfuscation techniques, including Base64 encoding and homoglyph substitution (e.g., іgnore), are present in SKILL.md and references/advanced-techniques.md. These strings serve as examples of attack vectors and do not contain hidden malicious code.
[INDIRECT_PROMPT_INJECTION]: The skill's core functionality requires ingesting and processing untrusted strings (prompt injections), creating a significant attack surface. Ingestion points include user-supplied prompts and model-generated attacks, while the agent's capabilities include network API calls and file writes. The skill lacks explicit internal sanitization for the agent's own handling of these external payloads.

prompt-clash