The Agent Skills Directory

[OBFUSCATION]: The skill documentation includes examples of URL bypass techniques using homoglyphs and zero-width characters. Specifically, references/oauth-attack-vectors.md provides an example of a homoglyph attack using a Cyrillic 'a' (\u0430) in https://tаrget.com and a zero-width space in https://target%E2%80%8B.com. These are explicitly described as security attack vectors for testing purposes.\n- [DATA_EXFILTRATION]: Proof-of-concept snippets in references/oauth-attack-vectors.md demonstrate how to exfiltrate intercepted OAuth codes and tokens to an external attacker-controlled domain (https://attacker.com/steal). This includes methods using the fetch API, CSS background images, and HTML form submissions.\n- [COMMAND_EXECUTION]: The skill provides Python code examples in references/oauth-attack-vectors.md that utilize the requests library to demonstrate token replay and authorization code reuse attacks. An agent utilizing this skill may be prompted to execute these scripts during a security engagement.\n- [INDIRECT_PROMPT_INJECTION]: The skill involves processing data from external OAuth endpoints which could be a vector for indirect injection from malicious targets.\n
Ingestion points: The skill targets metadata and configuration from endpoints like /.well-known/openid-configuration and /oauth/authorize as described in SKILL.md.\n
Boundary markers: No specific delimiters or safety instructions are defined for processing remote authentication metadata.\n
Capability inventory: The skill includes capabilities for network requests (Python requests), browser-side data exfiltration (fetch), and credential manipulation.\n
Sanitization: The instructions do not specify sanitization or validation of the remote service configuration data before processing.

oauth-sso-attack