sbh

Warn

Audited by Gen Agent Trust Hub on Nov 2, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPERSISTENCEPRIVILEGE_ESCALATIONDATA_EXFILTRATIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses various shell commands and system utilities like systemctl, jq, pgrep, which, find, and kill to manage the daemon and inspect system state.
  • [PERSISTENCE]: The sbh install command facilitates the installation of the tool as a persistent background service using systemd user units on Linux or launchd agents on macOS.
  • [PRIVILEGE_ESCALATION]: The documentation mentions the use of sudo for administrative tasks such as installing system-wide services and editing systemd unit overrides.
  • [DATA_EXFILTRATION]: The configuration supports a webhook notification system that allows the tool to send data to external URLs, which could be used to exfiltrate system information or activity logs.
  • [REMOTE_CODE_EXECUTION]: The sbh update and sbh install --from-source commands indicate functionality for downloading and executing or building code from external sources.
  • [INDIRECT_PROMPT_INJECTION]: The skill scans local directories for build artifacts. Untrusted data in the form of file names or directory paths could influence the agent's interpretation of cleanup reports and explainability logs.
  • Ingestion points: Filesystem scan via scanner.root_paths and artifact name patterns.
  • Boundary markers: None specified in the documentation for separating untrusted filenames from agent instructions.
  • Capability inventory: File deletion (rm), process termination (kill), and service management (systemctl).
  • Sanitization: Uses a Bayesian scoring engine to weigh factors before taking action, but does not explicitly detail input sanitization for filenames.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 2, 2026, 03:55 AM