Hamming AI — agentic threat model

Not certain from the listing — the underlying LLMs used to generate caller personas and simulate conversations are not specified. Threats include model reprogramming or adversarial prompt injection during simulated calls, which could cause the testing agent to output toxic content or fail to execute test plans.

Not certain from the listing — the storage of 'golden conversation sets' and test data is not detailed. Threats include data poisoning of these golden sets, which would corrupt baseline evaluations, or exfiltration of sensitive conversation logs.

The agent orchestrates multi-step simulated calls, generating dynamic caller personas and handling interruptions. Threats include insecure tool integration with voice platforms (VAPI, Retell) and potential manipulation of the test-generation logic to bypass safety checks.

Not certain from the listing — the hosting environment and sandboxing of simulated calls are not described. Threats include infrastructure compromise leading to the theft of API keys for integrated voice platforms (Synthflow, Bland, Retell).

This is the core layer of Hamming AI, providing real-time analytics, heartbeat monitoring, and red-teaming. Threats include evaluation gaming (where a compromised voice agent learns to pass Hamming's static 'golden sets') and blind spots in alert mechanisms (e.g., suppressing Slack/email alerts).

Not certain from the listing — no specific compliance certifications (e.g., SOC2, HIPAA) or access control mechanisms are mentioned despite handling healthcare and drive-thru voice data. Threats include unauthorized access to test suites and monitoring dashboards.

Highly relevant as Hamming directly interacts with other voice AI agents (Synthflow, Retell, etc.). Threats include cascading failures if a compromised target agent exploits Hamming's testing interface, or trust abuse where Hamming's simulated inputs are used to exploit vulnerabilities in the target agents.