5dive-cli

Fail

Audited by Snyk on Nov 16, 2026

Risk Level: CRITICAL
Full Analysis

CRITICAL E004: Prompt injection detected in skill instructions.

  • Potential prompt injection detected (medium risk: 0.30). The skill documents a legitimate 5dive CLI with disclosed, purpose-aligned behavior, but its gate rules steer agents to route decisions away from the human owner (calling human gates "a decision you find uncomfortable" and forbidding re-filing to reach a person), which weakens the user's control beyond what the task needs.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

  • Third-party content exposure detected (medium risk: 0.30). The required task-queue workflow (5dive task ls/show, and the heartbeat that wakes seats with queued rows) reads task titles and bodies, and 5dive trigger add github (--task-title) turns outsider-authored GitHub issue text into ordinary queue rows without the agent selecting a specific item, so this is a medium-risk path.

MEDIUM W009: Direct money access capability detected (payment gateways, crypto, banking).

  • Direct money access detected (high risk: 0.70). The skill mainly covers agent orchestration, task queues and secret gates, which are not financial operations. It also documents a payment-credential route. sudo 5dive secret link --card stores a card's number, expiry, CVC and holder name. A 5dive browser act step then types them into a site using "saved":"<name>.number|.exp|.cvc|.holder". That lets the agent make purchases or charges through a browser, which is a direct financial operation. No service-enforced authentication or approval is documented on the agent's use of the saved card. The owner supplies the card through a one-time link, but the agent then references the card by name. The "low-limit virtual card" caution is a spending limit, and the money-gate tier floors depend on the agent filing a task need gate, so neither counts as independent access control. I judge it a direct operation with no qualifying control.

MEDIUM W013: Attempt to modify system services in skill instructions.

  • Attempt to modify system services in skill instructions detected (medium risk: 0.55). The skill tells the agent to run sudo 5dive agent create, which by its own description creates a new Linux user (agent-<name>) and a systemd unit. It also covers writing /etc/5dive/connectors/tools.sh and similar root-level state. This is a vendor CLI with a scoped sudo grant and a teardown step (agent rm), and the skill never asks for NOPASSWD:ALL or a security bypass, which lowers the severity.

Issues (4)

E004
CRITICAL

Prompt injection detected in skill instructions.

W011
MEDIUM

Third-party content exposure detected (indirect prompt injection risk).

W009
MEDIUM

Direct money access capability detected (payment gateways, crypto, banking).

W013
MEDIUM

Attempt to modify system services in skill instructions.

Audit Metadata
Risk Level
CRITICAL
Analyzed
Nov 16, 2026, 11:13 PM
Issues
4