The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides numerous pre-crafted payloads in references/llm-attacks.md specifically designed to override an AI's existing instructions, such as 'Ignore all previous instructions' and role-playing scenarios like 'DAN'. These are intended to demonstrate how to bypass an agent's safety constraints.
[OBFUSCATION]: references/llm-attacks.md includes implementation of hide_in_zero_width using Unicode characters U+200B (Zero-Width Space) and U+200C (Zero-Width Non-Joiner) to encode instructions that are invisible to human users but processed by AI models. It also provides a unicode_evasion function in references/adversarial-ml.md that uses homoglyph substitution (replacing Latin characters with Cyrillic lookalikes) to evade text-based detection filters. Further techniques include Base64 encoding and HTML entity encoding for payload delivery.
[DYNAMIC_EXECUTION]: The skill uses torch.load() in SKILL.md and references/model-attacks.md to load model weights. This function uses pickle-based deserialization which is inherently unsafe and can result in arbitrary code execution when processing untrusted model files provided by external sources.
[INDIRECT_PROMPT_INJECTION]: The toolkit describes and implements techniques for 'Indirect Prompt Injection' in references/llm-attacks.md, showing how to embed instructions within documents or web pages that an AI might fetch or summarize, thereby influencing the agent's behavior through external data sources. It specifically notes how models fail to distinguish between instructions and data.

ctf-ai-ml