The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructions and example code demonstrate reading sensitive authentication data, specifically 'STUDIO_SESSION' and 'STUDIO_CSRF' tokens, from a 'studio.env' file. This sensitive information is then used as headers and cookies in 'curl' requests sent to the address defined in the '$STUDIO_URL' variable. While this is a standard mechanism for authenticated API interaction, it involves the retrieval and transmission of credentials.
[COMMAND_EXECUTION]: The skill utilizes several shell-level commands to perform its functions, including 'curl' for network communication, 'cat' for temporary file creation, and 'node' for executing a local authentication preflight script. These commands are executed within the agent's environment to interact with the external Studio API.
[DYNAMIC_EXECUTION]: The example workflow uses 'python3 -c' to execute a one-liner script that reads a local file and generates a JSON payload. This dynamic generation ensures that any user-provided HTML content is properly escaped before being submitted to the API, which serves as a security control against malformed JSON or shell injection.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection as it processes external, untrusted data. Ingestion points: The skill accepts user-provided HTML strings ('data') to be stored in edX blocks. Boundary markers: There are no explicit delimiters or specific instructions for the agent to ignore embedded commands within the provided HTML content. Capability inventory: The skill has access to network tools ('curl'), file system tools ('cat'), and language runtimes ('python3', 'node'). Sanitization: The risk is mitigated by the implementation of 'json.dumps()' in the Python execution step, which correctly sanitizes the input for transmission, preventing it from breaking out of the JSON structure or the shell command.

iblai-api-studio-html