The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze external web resources, which involves ingesting untrusted data such as HTTP headers, page source code (JavaScript), and SAML responses. This data could potentially be manipulated by an attacker to include malicious instructions.
Ingestion points: Target URLs, response headers, and client-side code (JavaScript) are all processed during the vulnerability hunting phases (SKILL.md).
Boundary markers: The skill explicitly defines a "Mandatory confirmation gate" that requires the user to provide the target scope and confirm each command before the agent transitions from read-only to active testing (SKILL.md).
Capability inventory: The skill instructs the agent to use tools like curl, subfinder, and custom Python scripts to interact with external targets (SKILL.md).
Sanitization: Security is primarily maintained through user oversight and the mandatory confirmation mechanism, which serves to validate all output generated from untrusted data before it is executed as a command.
[COMMAND_EXECUTION]: The skill provides templates for shell commands using standard networking and security tools (e.g., curl, grep, subfinder) to probe targets for authentication bypass vectors (SKILL.md). These commands are intended to be executed by the agent upon user approval.
[DYNAMIC_EXECUTION]: The skill includes Python script snippets for manipulating security tokens, such as stripping signatures from SAML XML or generating JWTs with the "none" algorithm (SKILL.md). These scripts are dynamically generated from templates and represent a standard but functional execution of code within the agent's context.

hunt-auth-bypass