grant-proposal
Warn
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill's optional 'Style reference' feature executes a bash script that takes a user-supplied
<source>argument from$ARGUMENTS. Specifically, it runspython3 "$STYLE_HELPER" --source "<source>". Direct interpolation of user-controlled strings into shell commands without explicit sanitization patterns creates a risk of command injection if the input contains shell metacharacters like backticks or subshells. - [DATA_EXFILTRATION]: The skill contains logic to check for and read
~/.claude/feishu.json. This file is used to send notifications to the Feishu platform and typically contains sensitive webhook URLs or API tokens. Accessing files in the user's home directory outside the immediate project scope increases the risk of credential exposure or unauthorized data transmission. - [INDIRECT_PROMPT_INJECTION]: The workflow ingests content from numerous external sources, including web search results, literature surveys, and paper styles extracted from arbitrary URLs or local files. This content is then used to steer the generation of the grant proposal and is provided to reviewer sub-agents. The instructions lack explicit boundary markers or sanitization logic to prevent malicious instructions embedded in these external sources from influencing the agent's behavior.
- [EXTERNAL_DOWNLOADS]: The skill uses
WebSearchandWebFetchto query databases like KAKEN and NSF Award Search. While intended for research, this capability allows the skill to retrieve data from arbitrary external domains, which contributes to the indirect prompt injection surface.
Audit Metadata