The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted paper abstracts and interpolates extracted keywords into AI art prompts for external tools like Midjourney and DALL-E. An attacker could potentially embed malicious instructions in an abstract to influence the content of the generated prompts.
Ingestion points: scripts/main.py via the --abstract command-line argument or stdin.
Boundary markers: None. The script processes input text through simple tokenization and keyword matching without protective delimiters.
Capability inventory: Generates structured text prompts for image generators.
Sanitization: None. No filtering of instructional language is performed on the input text.
[METADATA_POISONING]: The SKILL.md security checklist claims that input file paths are validated to prevent directory traversal (../), but the logic in scripts/main.py does not implement any validation or sanitization, using the open() function directly on provided paths.
[DATA_EXFILTRATION]: The skill script allows reading from and writing to arbitrary local file paths based on command-line arguments. Without path validation, this could be misused to access or modify files outside the intended workspace if the execution environment is not strictly sandboxed.

graphical-abstract-wizard