stelow-product-ads
Fail
Audited by Gen Agent Trust Hub on Jul 18, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The file
references/cli-tools/agent_browser.mdinstructs the use ofnpx -y @earendil-works/pi-agent-browser. This command downloads and executes code from an unverified NPM package at runtime, representing a high-risk execution pattern. - [EXTERNAL_DOWNLOADS]: Multiple tool definitions (
references/cli-tools/safe-change.md,references/cli-tools/codequality-review.md) recommend installing software from unverified GitHub repositories, such asgithub.com/PriNova/pi-agent-codebase-workflowsandgithub.com/cursor/plugins(note: while Cursor is a known service, it is not explicitly listed as a trusted vendor). - [EXTERNAL_DOWNLOADS]: The skill references several third-party NPM packages for core functionality, including
@tintinweb/pi-subagents,@juicesharp/rpiv-ask-user-question, and@tintinweb/pi-tasks, none of which originate from verified trusted organizations. - [COMMAND_EXECUTION]:
references/cli-tools/read-config.mdutilizesnode -eto execute inline JavaScript snippets. While used for configuration parsing, dynamic code execution via the shell is a sensitive pattern that can be exploited if input variables are not properly handled. - [DATA_EXFILTRATION]: The automated browser tool documented in
references/cli-tools/agent_browser.mdprovides capabilities to visit arbitrary URLs and capture screenshots/DOM snapshots. This could be used to exfiltrate sensitive local data if an agent is redirected to a malicious endpoint while having internal context. - [PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection due to the ingestion of untrusted data from the web.
- Ingestion points:
agent-browsertool (inreferences/cli-tools/agent_browser.md) reads live web content. - Boundary markers: Absent in the provided tool documentation.
- Capability inventory: The orchestrator suite includes full shell access (
bash), file writes (todo.md), and the ability to spawn autonomous subagents (subagents.md). - Sanitization: No evidence of sanitization or instruction-filtering for content retrieved from external URLs.
Recommendations
- AI detected serious security threats
Audit Metadata