The Agent Skills Directory

[OBFUSCATION]: The file references/mcp-attack-payloads.md contains Base64 encoded instructions that, when decoded, direct the agent to exfiltrate .env files to an external server.
[OBFUSCATION]: The skill demonstrates hiding malicious instructions using ANSI escape sequences (\x1b[8m) to make text invisible in terminal outputs while remaining readable by AI models.
[OBFUSCATION]: Provides Python code and methodology for using zero-width Unicode characters (U+200B, etc.) and Unicode Tag characters to embed hidden instructions in tool descriptions.
[DATA_EXFILTRATION]: Payload examples specifically target highly sensitive files including ~/.ssh/id_rsa, ~/.env, and /etc/passwd, with instructions to send them to attacker.com domains.
[PROMPT_INJECTION]: Contains instructions designed to bypass agent constraints using high-priority delimiters such as <IMPORTANT>, <SYSTEM>, and [INST] within tool metadata.
[DATA_EXFILTRATION]: Describes 'Shadow Tool' attacks where a malicious MCP server's tool description hijacks the behavior of other trusted tools to redirect data (e.g., changing email recipients).

mcp-security