Vibe Kanban — agentic threat model

Not certain from the listing — relies on external foundation models (Claude, Gemini, Codex) driven via their respective CLIs; vulnerable to prompt injection or model reprogramming that could cause the underlying models to generate malicious code or bypass safety guardrails.

Not certain from the listing — operates on codebase files and kanban issue descriptions rather than a dedicated vector database; risk involves unauthorized data exfiltration of proprietary source code or poisoning of the codebase context provided to the agents.

The orchestration framework wraps and drives external agent CLIs (Claude Code, Gemini CLI, etc.) to execute tasks. Vulnerabilities include insecure tool integration, command injection via issue descriptions, and the risk of agents executing unintended shell commands or modifying files outside the intended scope.

Runs coding agents in isolated workspaces to execute tasks. This sandboxing is a critical control to prevent container escape, host compromise, and lateral movement, though the security depends heavily on the robustness of the isolation mechanism used for these workspaces.

Provides diff reviews with inline comments, enabling human-in-the-loop verification of agent outputs before merging. However, there is a risk of blind spots if malicious code changes are subtle or if reviewers suffer from automation bias and approve malicious diffs.

Not certain from the listing — open-source tool coordinating local or cloud CLI runs; lacks explicit mention of enterprise identity management, role-based access control (RBAC), or compliance auditing for who can dispatch agents or approve diffs.

Coordinates multiple concurrent agent runs (Claude Code, Codex, Gemini CLI, Amp) against a single kanban board. This multi-agent orchestration introduces risks of cascading failures, race conditions on shared codebases, and cross-agent trust abuse where one compromised agent influences another.