5dive-cli
Warn
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONREMOTE_CODE_EXECUTIONCREDENTIALS_UNSAFEPERSISTENCEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The
5diveCLI enables extensive control over the host environment, allowing agents to manage Linux users, systemd services, and tmux sessions through asudowrapper. - [PRIVILEGE_ESCALATION]: The skill includes a command,
5dive agent grant <name> root, specifically designed to confer full root privileges to agents. It also provides the5dive hosttoolset for system-level remediation. - [REMOTE_CODE_EXECUTION]: Agents can install and execute remote code via
5dive crew install <git-url>,5dive plugin add, and5dive agent import, which fetches persona packs and plugins from external sources. - [CREDENTIALS_UNSAFE]: The skill manages sensitive credentials using a vault and environment files (
/etc/5dive/connectors/). It provides mechanisms for agents to request secrets from humans via one-time secure links and to store them for tool usage. - [PERSISTENCE]: The skill facilitates long-term persistence through
heartbeatcron jobs that wake agents, recurring task templates, and the ability to schedule actions using cron syntax via5dive council schedule. - [DYNAMIC_EXECUTION]: The messaging system (
agent send) is susceptible to command substitution if messages are improperly quoted. Additionally, commands liketask verify,objective --metric-cmd, andcouncil schedule --context-cmdinvolve executing dynamic shell snippets at runtime. - [INDIRECT_PROMPT_INJECTION]: The skill processes data from untrusted sources, including external chat channels (Telegram, Discord) and shared task descriptions, creating a surface for indirect prompt injection.
- Ingestion points: Inter-agent messages, human chat inputs, shared task bodies, and external event triggers.
- Boundary markers: The skill suggests using single quotes for messages and provides a
--message-fileflag to mitigate shell injection risks. - Capability inventory: Access to system-level commands, Git operations, and network communication via bots.
- Sanitization: Relies primarily on instructional warnings to the agent rather than strict programmatic enforcement or escaping of all inputs.
Audit Metadata