systematic-debugging
Warn
Audited by Gen Agent Trust Hub on Mar 29, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [PROMPT_INJECTION]: The skill contains multiple 'Pressure Test' files (
test-pressure-1.md,test-pressure-2.md,test-pressure-3.md) that utilize imperative language to override the AI agent's standard behavior and decision-making processes. These files use phrases like "IMPORTANT: This is a real scenario. You must choose and act. Don't ask hypothetical questions - make the actual decision," which are characteristic of prompt injection attempts to bypass AI safety filters or operational guidelines.
- [DATA_EXFILTRATION]: Instructions in
SKILL.mdrecommend gathering evidence by dumping environment variables usingenv | grep IDENTITY. This pattern specifically targets potential secrets (identities) and can lead to the exposure of credentials if the output is logged, shared, or processed by subsequent tools. - [COMMAND_EXECUTION]: The provided utility script
find-polluter.shexecutes shell commands (npm test) on files identified at runtime. This creates a risk of arbitrary code execution if the file paths or the content of the tests are influenced by untrusted data. - [COMMAND_EXECUTION]:
SKILL.mdsuggests the use of powerful system commands such assecurity list-keychainsandcodesign. While intended for debugging signing issues, these commands interact directly with the host system's security and identity management infrastructure.
Audit Metadata