Kubiya — agentic threat model

Not certain from the listing — the specific foundation models used by Kubiya are not disclosed. The primary threat at this layer is adversarial prompt injection via Slack, which could trick the model into executing unauthorized DevOps commands.

Not certain from the listing — details regarding vector databases, RAG pipelines, or training data operations are not provided. The main threat is the potential exfiltration or poisoning of configuration metadata and codebase context used to inform the agent.

Kubiya translates natural language into complex DevOps actions and workflows. Framework-level threats include tool execution hijacking, where malicious inputs manipulate the orchestration layer to run destructive commands (e.g., deleting cloud resources) instead of intended tasks.

Not certain from the listing — the hosting architecture and sandboxing mechanisms for executing workflows are not detailed. A compromise here could lead to container escape or lateral movement into connected customer cloud environments.

Kubiya features built-in guardrails, templates, and audit trails for all actions. The primary threat is guardrail evasion through sophisticated prompt engineering, or logging bypasses that hide unauthorized infrastructure changes.

Kubiya implements Just-in-Time (JIT) permissions management and access controls. Threats include privilege escalation vulnerabilities within the JIT system, allowing unauthorized Slack users to execute highly privileged cloud operations.

The platform supports 'AI teammates' and customizable workflows. Threats include cascading failures or trust abuse if one compromised agent or workflow triggers unauthorized actions across other integrated DevOps tools and agents.