Duckie Agent — agentic threat model

Not certain from the listing — The underlying foundation models are proprietary and undisclosed. The primary threat is prompt injection from external customers interacting with the auto-responder, potentially leading to model reprogramming or leaking system instructions.

Not certain from the listing — The agent ingests customer support data and system logs for 'continuous learning' and 'log investigation'. This introduces severe risks of data poisoning (if malicious tickets are ingested) and data exfiltration of sensitive PII or secrets contained within logs.

Not certain from the listing — The orchestration framework is closed-source. Threats include tool misuse where the agent, triggered by a malicious prompt, performs unauthorized actions in Zendesk, Intercom, or Slack, or accesses restricted log files.

Not certain from the listing — Hosting and sandboxing details are not provided. Compromise of the deployment infrastructure could expose highly sensitive API keys and tokens used to integrate with Slack, Zendesk, and internal log databases.

Not certain from the listing — There is no mention of guardrails, output filtering, or evaluation frameworks. A lack of real-time monitoring could allow toxic, inaccurate, or malicious auto-responses to reach customers undetected.

Not certain from the listing — No compliance certifications (such as SOC2 or ISO 27001) or explicit access control mechanisms are detailed, raising concerns about how customer data privacy and regulatory alignment are enforced.

Not certain from the listing — While not explicitly a multi-agent system, the agent operates within a broader ecosystem of third-party platforms (Slack, Zendesk, Intercom), making it vulnerable to cascading failures or trust abuse if those connected platforms are compromised.