CatDoes — agentic threat model

Not certain from the listing — The specific foundation models powering CatDoes are not disclosed. Standard risks of LLM reprogramming, prompt injection, and adversarial inputs apply, particularly those that could manipulate the agent into generating insecure code.

The agent imports data directly from GitHub repositories and manages built-in backend services. This introduces risks of repository data exfiltration, exposure of sensitive environment variables, and potential poisoning of the codebase context during import.

The agent autonomously plans work, writes code, runs tests, and self-corrects errors. Vulnerabilities in this orchestration layer could allow an attacker to hijack the tool-calling mechanism, leading to unauthorized file modifications or execution of malicious commands during the test phase.

Because the agent executes code and runs tests in a cloud environment, robust container sandboxing is critical. A compromise at this layer could lead to container escape, lateral movement within the hosting infrastructure, or unauthorized access to the built-in backend services.

While the platform includes built-in error monitoring for the generated applications, the listing does not specify any LLM-specific guardrails, safety alignment checks, or anomaly detection to prevent the generation of malicious or vulnerable code.

Not certain from the listing — There is no mention of enterprise security compliance standards (e.g., SOC2, ISO 27001), role-based access control (RBAC), or audit logging for the autonomous actions taken by the cloud agent.

The agent integrates with external ecosystems by importing from GitHub and preparing builds for app stores. Compromising the agent's credentials or session tokens could allow unauthorized commits to GitHub or the injection of malicious updates into published mobile and web apps.