lov-png2svg

Warn

Audited by Gen Agent Trust Hub on Nov 17, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPERSISTENCE
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructions provide shell command templates that directly interpolate user-provided variables (INPUT_PNG, OUTPUT_SVG) into command strings without specifying sanitization or escaping requirements.
  • Evidence: In SKILL.md, the agent is instructed to execute commands such as magick INPUT_PNG -fuzz 15% ... INPUT_PNG.temp.png and vtracer --input INPUT_PNG.temp.png --output OUTPUT_SVG ....
  • Risk: If a user provides a file path containing shell metacharacters (e.g., ; rm -rf /, `id`), it could lead to arbitrary command execution on the host system if the agent platform does not enforce automatic shell escaping.
  • [PERSISTENCE]: The "General feedback loop" (通用反馈闭环) section in SKILL.md instructs the agent to modify its own "canonical Skill" files and "shared specifications" when user feedback is deemed "reusable".
  • Evidence: The instructions state: "reusable 先确定作用域:頖域规则先更新对应 canonical Skill;适用于所有 Skill 的规则先更新共享规范。"
  • Risk: This capability allows an attacker to use indirect prompt injection (via "feedback") to persist malicious instructions into the skill's source code, which would then affect all future sessions using that skill. Although the skill includes a human-in-the-loop checkpoint, the instruction to automate the modification of its own logic is a known persistence pattern.
  • [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its interaction with untrusted file paths and its high-privilege capabilities.
  • Ingestion points: File paths provided as INPUT_PNG in SKILL.md.
  • Boundary markers: Absent. There are no instructions for the agent to treat the input as untrusted or to ignore instructions embedded in metadata.
  • Capability inventory: The skill utilizes magick, vtracer, and npx svgo, all executed via shell subprocesses with file-system write and delete (rm) permissions.
  • Sanitization: Absent. The instructions do not specify validation or sanitization for the interpolated strings.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 17, 2026, 07:05 AM