The Agent Skills Directory

[PROMPT_INJECTION]: The agent instructions include specific defenses against indirect prompt injection. Reviewers are directed to treat manuscript content as read-only data and are explicitly told to ignore any imperative commands embedded within that data. The static detector flags for 'ignore instructions' are false positives triggered by these defensive markers.
[DATA_EXFILTRATION]: The skill maintains a narrow scope of operation, focusing on reading manuscripts and generating reports. There is no evidence of unauthorized file access to sensitive directories or exfiltration of data to external domains.
[COMMAND_EXECUTION]: The workflow involves local integrity checks through scripts like 'check_pipeline_integrity.py'. These are internal consistency checks within a controlled pipeline and do not expose the system to arbitrary command injection from external sources.

academic-paper-reviewer