The Agent Skills Directory

[PROMPT_INJECTION]: The skill's instructions are strictly professional and domain-specific. No patterns were found attempting to bypass safety guardrails, override system instructions, or extract prompt history.
[DATA_EXFILTRATION]: The core logic in scripts/main.py is limited to local file system I/O (reading references and writing screening results). There are no network-capable libraries (like requests or urllib) or commands (like curl or wget) used to transmit data externally.
[REMOTE_CODE_EXECUTION]: All code is contained within the local scripts/ directory. The skill does not perform any remote script downloads, piped shell executions, or dynamic fetching of executable content.
[CREDENTIALS_UNSAFE]: A thorough scan of SKILL.md, scripts/main.py, and references/criteria_template.yaml revealed no hardcoded API keys, secrets, or authentication tokens.
[COMMAND_EXECUTION]: While the skill provides examples for running the screener via the command line (python scripts/main.py), it does not include instructions for the agent to execute arbitrary or privileged shell commands. The usage is restricted to the provided Python script.
[OBFUSCATION]: All instructions and code files are provided in plain text. No Base64, hex encoding, zero-width characters, or hidden acrostics were detected.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data (academic abstracts). However, the processing is handled by a local Python script using regex and string matching rather than interpolating the data directly into an LLM prompt, effectively neutralizing the risk of data-driven instruction injection.

systematic-review-screener