The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses sudo 5dive commands to perform administrative tasks. It explicitly warns about a command injection vulnerability in the agent send command: when using the --message= flag with double quotes, shell command substitution (e.g., `command` or $(command)) will execute as the calling user. The documentation recommends using --message-file or single quotes to mitigate this risk.
[PRIVILEGE_ESCALATION]: The skill provides administrative commands to elevate agent permissions. The 5dive agent grant <name> root command confers unrestricted root access (the ability to run any command as any user) on an agent. Agents can also be created with an admin isolation tier, which provides a scoped sudo grant for 5dive and systemd services.
[REMOTE_CODE_EXECUTION]: The 5dive crew install <git-url> command allows for the cloning and installation of code from arbitrary external Git repositories into a local virtual environment. This code can then be executed via 5dive crew run.
[EXTERNAL_DOWNLOADS]: Several commands facilitate fetching content from remote sources. 5dive agent install downloads agent binaries, 5dive plugin add fetches extensions, and 5dive market allows agents to be hired from a marketplace. These downloads typically target vendor-controlled or marketplace infrastructure.
[PERSISTENCE]: The skill manages persistent system resources. Agents are deployed as systemd units and tmux sessions that persist across reboots. The 5dive council schedule command installs managed crontab entries, and the 5dive trigger system sets up signed listeners for external events like GitHub webhooks.
[INDIRECT_PROMPT_INJECTION]: The agent environment has a significant surface for indirect prompt injection. It ingests untrusted data from external chat channels (Telegram, Discord) and GitHub webhooks. This content is passed to agents as input, potentially allowing external actors to influence agent behavior. Boundary markers are limited to instructional guidance regarding shell quoting.
[DATA_EXFILTRATION]: Capabilities for moving data off-host are present. 5dive agent export packages agent state and memory for export. 5dive push and 5dive deploy interact with external services to transmit code. 5dive memory search allows for the retrieval of team decisions and research data.
[DYNAMIC_EXECUTION]: The 5dive acp command dynamically stages and runs a TypeScript server using the bun runtime. The 5dive loop spawn command allows agents to dynamically generate and execute new tasks, creating a chain of execution.

5dive-cli