A Nitro Enclave is a locked-down VM carved from a parent EC2 instance: no persistent storage, no interactive SSH, no operator console, connectivity primarily over vsock. The parent cannot dump the enclave's memory with ordinary tools. You use it when the threat model includes the instance role's humans, a compromised parent OS, or a need to prove to KMS that only a measured enclave image can unwrap a key.

Attestation is the product

At boot, Nitro produces an attestation document with PCRs (hashes of the enclave image, IAM role, and related measurements). KMS can require those PCRs in a key policy (kms:RecipientAttestation:ImageSha384 and friends) so Decrypt only works from inside an enclave whose image matches. The parent can broker bytes over vsock; it should never see plaintext keys.

text
parent  --vsock--> enclave (unwrap DEK via KMS + attestation)
enclave processes payload, returns ciphertext or a token

If you skip attestation-bound KMS and just "run the code in an enclave," you have isolation from the parent but not a cryptographic proof. Anyone who can start a different enclave image on a similar instance may still call KMS if the policy is slack.

What you will fight operationally

Images are built with nitro-cli. Debugging is log-over-vsock, not ssh. Memory and vCPU are reserved from the parent; oversize the enclave and the parent starves. No easy outbound networking: you proxy through the parent, which means the parent is still in the availability path and can DoS or tamper with ciphertext (integrity of the channel is your protocol). Clock and entropy come from Nitro; still design protocols that do not need the parent to be honest about time.

Enclaves are a poor fit for a general application server. They are a fit for: pinning a TLS private key, processing PCI tokens, or a small policy engine. Latency includes enclave boot if you start per job; long-lived enclaves are more common.

Honest limits

Nitro Enclaves are not a substitute for IAM hygiene. A parent that can replace the enclave image in an auto-scaling group without PCR pinning in KMS undoes the story. Side channels exist in the confidential computing literature; AWS's isolation is strong against casual operator dump, not a mathematical end of all leakage.

Read the user guide's attestation and KMS integration chapters. Then write a key policy that names PCRs before you write the vsock protocol. The enclave is a vault. The policy is the lock.