MetaMask has taken a decisive step into a new era of autonomous finance, announcing the public launch of its non-custodial Agent Wallet. This event, which I mentioned in my previous reviews as an inevitable stage in the evolution of Web3, marks a transition from simply storing assets to delegating their management to artificial intelligence.

Agent Wallet operates through a command-line interface (CLI) and is integrated with leading AI platforms, including Claude Code, Codex, OpenClaw, Hermes, OpenCode, and Cursor. The core of the innovation is that the user sets clear parameters — spending limits, a list of approved protocols, and other rules — and the AI agent then independently executes on-chain operations without exceeding these boundaries.

A key security element is Guard Mode, activated by default. It includes allowlists of networks, addresses, and token recipients, as well as a daily limit on outgoing funds. The developers call this mechanism two-factor authentication (2FA). If an operation falls outside the established rules or seems suspicious to the system, it is paused. The user can approve or reject the transaction via the MetaMask mobile app or by email, with the request automatically expiring after five minutes without a response.

For those ready to trust AI to a greater extent, Beast Mode is provided. In this mode, allowlists and daily limits are disabled, but checks for malicious operations remain. This is a compromise between autonomy and security, which, in my view, still requires a high degree of awareness from the user.

Key Isolation and a New Security Model

Architecturally, Agent Wallet solves the main problem of trust in AI. In server mode, private keys are stored in a trusted execution environment (TEE), isolated from the AI agent process. This means the agent only transmits a request for an operation, while separate infrastructure verifies compliance with policies and signs the transaction. The user retains control, with the ability to export the seed phrase, which allows the wallet to be called non-custodial, albeit with the caveat of dependence on MetaMask infrastructure for daily operations.

Financial Protection and Unavoidable Risks

MetaMask also introduced the Transaction Protection program. It offers compensation of up to $10,000 per month if a transaction deemed safe still results in covered damage. Payments are made in the mUSD stablecoin from Bridge (owned by Stripe) based on the M0 protocol. However, it is important to understand: coverage does not extend to key compromise, market losses, or protocol exploits.

The main risk I see is prompt injection. A malicious instruction can be hidden in a web page, token description, or API response. MetaMask attempts to mitigate this threat by separating the functions of proposal and execution, but in Beast Mode, where the agent is not restricted by lists, a correctly appearing transaction caused by manipulation may still pass checks. This is a fundamental challenge for the entire industry, and no perfect solution exists yet.

The launch of Agent Wallet is a landmark event. Coinbase has already presented similar solutions, and Base has launched MCP for on-chain access. We are witnessing the formation of new infrastructure where AI becomes a full-fledged participant in financial markets. This opens up enormous opportunities, but requires us, as a community, to develop new standards of security and responsibility.