whiyh-gpt-hill-climb
Fail
Audited by Gen Agent Trust Hub on Jun 14, 2026
Risk Level: HIGHCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill frequently executes shell commands and npm scripts (e.g.,
npm run telemetry,npm run dev,npm test,scripts/check-fast.sh) to interact with the repository, collect data, and validate changes. - [REMOTE_CODE_EXECUTION]: The skill operates in a self-modifying loop where it generates new source code—including prompts, logic, and routing—based on telemetry and then executes that code via the development server and test suites. This provides a mechanism for external input to drive code execution on the local filesystem.
- [PROMPT_INJECTION]: The skill ingests telemetry data, such as transcripts of user gameplay, and uses this information to drive automated code changes. This exposes the agent to indirect prompt injection, where malicious input embedded in the telemetry could influence the agent's reasoning or the code it produces.
- [DATA_EXFILTRATION]: The skill reads telemetry data and transcripts that may contain sensitive user interactions. While the skill instructs the agent not to expose this data publicly, the broad capability to access and process this information is a significant data exposure risk.
Recommendations
- AI detected serious security threats
Audit Metadata