AgentOps — agentic threat model

Not certain from the listing — AgentOps is model-agnostic and integrates with various LLMs. It does not host foundation models itself, but its telemetry captures model inputs/outputs, making it a target for harvesting proprietary prompts or model responses.

AgentOps collects and stores telemetry data, logs, and saved completions for fine-tuning. Threats include data exfiltration of sensitive prompt/completion history and potential poisoning of fine-tuning datasets stored or processed via the platform.

Integrates directly with frameworks like CrewAI, Langchain, and Autogen by hooking into tool calls and agent execution. Vulnerabilities in the SDK could allow an attacker to bypass logging or exploit the host framework via malicious telemetry payloads.

Not certain from the listing — AgentOps is a Python SDK sending data to a cloud platform (or self-hosted). Infrastructure threats include insecure transmission of API keys/telemetry, and unauthorized access to the AgentOps dashboard.

This is AgentOps' primary layer. It provides debugging, auditing, cost tracking, and prompt injection detection. Threats include blind spots if the SDK is bypassed, tampering with audit logs, or evasion of its prompt injection detection mechanisms.

Not certain from the listing — The listing mentions 'Debug and Audit' and tracking prompt injections, but does not specify compliance standards (like SOC2, GDPR) or RBAC controls for the dashboard.

Monitors multi-agent interactions. If compromised, it could act as a single point of failure or a side-channel attack vector to spy on complex multi-agent workflows across an enterprise.