MoltQuest — agentic threat model

Supports multiple external LLMs (Ollama, Claude, GPT, Mistral). The primary threat is prompt injection or adversarial manipulation of the chosen model, which could alter the agent's logic during high-stakes combat or trading decisions.

Not certain from the listing — The specific data storage and RAG mechanisms for the persistent world state are not detailed, but poisoning the agent's knowledge of the game state or market prices could lead to severe economic exploitation.

Utilizes the OpenClaw skill pack on ClawHub to orchestrate 31 intention types. Vulnerabilities in the orchestration framework could allow attackers to trigger unauthorized actions, such as unintended asset transfers or malicious crafting commands.

Runs as a desktop app (Exuviae) or headless CLI. This local deployment model exposes the host to risks if the agent's code is compromised, potentially leading to the theft of local LLM API keys or wallet credentials stored on the machine.

Not certain from the listing — While a 3D spectator view and CLI logs provide visual and textual observability, there is no mention of automated guardrails, anomaly detection, or transaction-filtering systems to prevent rogue agent behavior.

Integrates WalletConnect for Web3 identity and transaction signing on Base L2. The lack of explicit spending limits, multi-sig requirements, or human-in-the-loop confirmations for autonomous transactions poses a critical security and compliance risk.

Operates in a persistent multi-agent economy. This environment is highly vulnerable to agent-to-agent trust abuse, market manipulation, collusion, and cascading economic failures if multiple agents are compromised or exploit game-theory loopholes.