Diffblue Cover — agentic threat model

Not certain from the listing — The specific foundation models or proprietary reinforcement learning engines used are not disclosed. Potential threats include model evasion or manipulation of the RL reward function if an attacker can feed malicious code patterns to the agent.

Operates directly on local Java source code. The primary threat is source code exposure or data exfiltration, which is heavily mitigated by the agent's on-premise deployment model that keeps intellectual property within the enterprise boundary.

The agent orchestrates test generation and execution. A key threat is tool misuse or manipulation where the agent is tricked into generating tests that execute malicious payloads, or exploiting vulnerabilities in the underlying test execution framework.

Runs on-premise and explicitly utilizes sandboxed execution to run the generated tests. This sandboxing is critical to prevent container escape, privilege escalation, or lateral movement during the execution of untrusted or newly generated test code.

Not certain from the listing — While the agent ensures deterministic and compile-correct outputs, the listing does not detail the internal logging, telemetry, or guardrails used to monitor the reinforcement learning loop for drift or anomalous behavior.

Not certain from the listing — On-premise deployment aligns with strict IP security policies, but specific compliance certifications (such as SOC 2, ISO 27001) or detailed access control policies are not explicitly mentioned.

Integrates directly into developer IDEs (like IntelliJ) and CI/CD pipelines. The primary threat is pipeline compromise, where a compromised agent could push malicious test code that gets automatically executed in trusted build environments.