2026-07-04 · ledger
Ledger published a security-focused post arguing that AI agents should not be trusted with direct control of keys. The company frames Ledger Agent Stack as a separation-of-duties model: agents can search, plan, and propose, but humans approve risky actions on secure Ledger hardware.
Ledger described a model where identity, policy, and hardware approval work together to keep agents useful without giving them unchecked authority. The core message is that agent identity alone is not security; high-risk actions still need transaction-time approval.
If an AI agent can directly move assets, disclose secrets, change permissions, or sign transactions, one compromised instruction path can become a full custody failure. The damage is not limited to one bad prompt; it can include irreversible transfers, exposed recovery data, and polluted audit trails.
Cold storage keeps private keys and recovery material outside the reach of autonomous software. Hardware approval creates a final human-controlled checkpoint, so agents can prepare actions without ever holding the keys needed to execute them.
Read Original Post →