AutoGPT — agentic threat model

Uses GPT-4 as its foundation model. Highly vulnerable to prompt injection, goal hijacking, and indirect prompt injection via web-browsing, which can redirect its autonomous execution loops.

Not certain from the listing — the description mentions internet access for information gathering but does not detail its vector database, memory storage, or RAG architecture, leaving it vulnerable to data poisoning via untrusted web content.

As an autonomous framework, it orchestrates planning, memory, and tool calling. It possesses high risk of tool misuse and insecure tool integration, particularly because it can autonomously write and execute code to achieve goals.

Not certain from the listing — the deployment environment and sandboxing controls for executing the generated code are not specified, posing extreme risks of host compromise and privilege escalation if run unsandboxed.

Not certain from the listing — there is no mention of built-in guardrails, real-time monitoring, or evaluation frameworks to detect, log, or intercept malicious or unintended autonomous actions.

Not certain from the listing — no identity, authorization, policy enforcement, or compliance governance mechanisms are described for managing agent permissions.

Not certain from the listing — while it acts as an autonomous agent, the listing does not detail multi-agent coordination or marketplace interactions, though cascading failures remain a risk if it spawns sub-agents.