The Agent Skills Directory

[PROMPT_INJECTION]: No patterns of prompt injection, jailbreaking, or instructions to ignore safety guidelines were detected in the skill instructions.
[DATA_EXFILTRATION]: The skill focuses on local filesystem operations within the paper directory. There are no indications of unauthorized data transmission or access to sensitive credentials.
[COMMAND_EXECUTION]: The skill uses Bash for recompiling LaTeX source and managing state. These commands are localized to the project environment and align with the skill's purpose.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill ingests untrusted content from LaTeX source files (.tex) during the review process.
Boundary markers: Security is enhanced by the 'Reviewer Independence Protocol,' which executes reviews in isolated, ephemeral sessions to prevent context contamination.
Capability inventory: The skill is limited to Read, Write, Edit, Bash, and Glob tools.
Sanitization: The 'edit-whitelist' provides a policy-based constraint on modifications, ensuring that potentially malicious instructions in reviews cannot affect frozen document components.

auto-paper-improvement-loop