The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data through the $ARGUMENTS variable, which is parsed for error messages, reproduction steps, and failure patterns. This creates an attack surface where an attacker could embed malicious instructions within logs or reproduction steps to influence agent behavior.
Ingestion points: The $ARGUMENTS variable is used in the 'Context' section to process issues and again at the end of the skill instructions.
Boundary markers: The skill does not define clear delimiters (like triple backticks or tags) or explicit 'ignore embedded instructions' warnings for the input data.
Capability inventory: The agent is instructed to query observability platforms (Sentry, DataDog, etc.), generate code fixes, and execute validation tests.
Sanitization: While the workflow suggests manual review and authorization, there are no mechanical sanitization steps for the input data strings.
[DYNAMIC_EXECUTION]: The workflow directs the agent to generate code fixes and execute test suites based on the diagnosis of the provided issue. This represents a dynamic execution surface where code generated at runtime (the fix) is subsequently executed during the validation phase. If the code generation process is successfully manipulated by malicious content in the input report, the agent could potentially generate and execute harmful code during validation.

error-diagnostics-smart-debug