share
Warn
Audited by Gen Agent Trust Hub on May 11, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONCREDENTIALS_UNSAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill's primary workflow involves bundling session history from
~/.claude/projects/and uploading it to an external Cloudflare R2 bucket usingaws s3 cp. This represents the intentional exfiltration of sensitive development logs, conversation history, and potential environment metadata. - [CREDENTIALS_UNSAFE]: The execution plan relies on the 1Password CLI (
op) to retrieve credentials for the Cloudflare R2 upload, and the use ofaws s3 presignto generate publicly accessible URLs for the uploaded data. - [COMMAND_EXECUTION]: The instructions direct the agent to execute multiple local shell scripts (
scripts/bundle.sh,scripts/sanitize.sh) and system utilities includingbrotli,aws,op, andbun. - [PROMPT_INJECTION]: The skill contains a 'Self-Evolving Skill' section that explicitly commands the agent to 'fix this file immediately' and 'update this file' based on its execution experience. This instruction facilitates self-modification of the agent's own behavior and rules, which can be exploited to persist malicious logic or bypass future constraints.
Audit Metadata