The era of 'can they act?' is over. As of November 2026, the critical question for autonomous agents is how to keep them safe, accountable, and recoverable when they inevitably crash or get spoofed. With Shopify reporting a tripled year-over-year increase in AI-driven traffic in Q2 2026, and analysts projecting up to $500 billion in agent-mediated US e-commerce by 2030, the security model that assumes a human at the keyboard is broken. Agents now ingest untrusted content and execute tool calls autonomously, requiring a security architecture designed for machines, not just humans.
Identity: Beyond Self-Attestation
Traditional identity strings in request headers are useless for agents—they are merely self-attestation. The new standard, as outlined by NIST in their May 2026 principles and implemented in protocols like SAAX, requires cryptographic identity. An agent’s proofs must be signed with a key bound to a principal, using standards like EIP-712 for domain-separated messages. Crucially, agent identity is distinct from the owner’s identity; it represents delegation. This ensures that Agent A and Agent C can act for the same principal without being interchangeable, preventing identity spoofing where a counterparty claims another agent’s identity by manipulating caller-supplied arguments.
Authorization: Scoped, Bounded, and Revocable
Knowing who is acting is only half the battle; controlling what they can do is the other. Least privilege is no longer a suggestion but a requirement, with permissions scoped to specific commitments and resources rather than global grants. The SAAX implementation demonstrates this through a durable authorization registry that binds agent IDs to roles and action allowlists with strict validity windows. If a grant expires, it fails closed. If revoked, the change is recorded in an append-only log, ensuring that even replayed operations are rejected. This separation of roles prevents scenarios where an evaluator submits evidence as a seller or an agent declares its own evidence verified.
Recovery: Surviving the Crash
The hardest problem in agent security isn't identity or permission—it's what happens when the process dies mid-transaction. The solution lies in durable state and reconciliation before mutation. Commitment records must live in persistent storage, independent of the agent’s memory. If an agent terminates, a replacement agent can resume the workflow only if it has its own distinct authorization. The SAAX protocol demonstrates this with idempotent transitions and version enforcement, ensuring that retries don't double-charge and concurrent agents don't silently overwrite each other. In a recent nonfinancial beta, a replacement agent successfully resumed a commitment after the original buyer agent terminated, with the event journal clearly recording the replacement’s identity.
Key Takeaways
- NIST published ten principles for AI agent identity and authorization in May 2026, emphasizing delegated access and audit trails.
- Cryptographic signatures (e.g., EIP-712) are replacing self-asserted headers as the basis for agent identity.
- Authorization must be commitment-scoped, expiring, and revocable, with changes taking effect immediately even against replayed requests.
- Recovery relies on durable state, independent authority for replacement agents, and idempotency to prevent duplicate side effects.
- The market is moving fast: Bain projects $300–500B in agent-mediated e-commerce by 2030, making these controls critical for infrastructure.
The Bottom Line
Agent security is not a single control but a triad of identity, authorization, and recovery. If you cannot cryptographically verify who is acting, bound their permissions to specific tasks, and seamlessly recover from crashes, your autonomous workflows are just expensive bugs waiting to happen.