
Zero Operator Access on AWS: How Bee Locks Out Its Own Engineers
Cloud Platform EngineeringBee, an Amazon company that makes a wearable AI device, listens to conversations about users’ health, finances and family. It is built so that Bee engineers, other Amazon employees and AWS infrastructure operators all have no way to read that data. The surprise is how little of it rests on policy. For cloud teams handling sensitive data, Bee’s design shows what “we can’t see your data” looks like when a user’s phone can check it for itself.
How It Works
AWS’s write-up of Bee’s architecture calls the design Zero Operator Access (ZOA). The core control is EC2 instance attestation. A user’s device releases decryption keys only to instances running software it has verified.
Measured boot. Bee builds attestable AMIs with no SSH, SSM or shell, a read-only root filesystem, and local storage encrypted with an in-memory key. Instances are replaced rather than rebooted. Reference values for PCR4, PCR7 and PCR12 are computed offline, before any instance launches. At boot, measurements extend into NitroTPM, and the AWS Nitro System signs the attestation document.
Covering the workload. A clean OS isn’t enough if someone can swap a container or model file. A custom tool resolves every container to its immutable ECR digest, hashes each S3 artifact, and extends PCR14 with one value across the manifest. Any change to an image or model weight changes the attestation.
Independent transparency. A separate Amazon team that Bee doesn’t control recomputes the expected PCRs and commits them to Sigstore Rekor, a public transparency log. An instance checks itself against that log, then gets a TLS certificate from AWS Private CA with the attestation and proof embedded. The Bee app verifies it offline against the Nitro root and refuses to connect if anything changed.
Keys and network. Only then does the app release a short-lived key. It lives only in memory and in Redis stores with persistence disabled, expires after 7 days without user activity, and is never held by AWS. Data reaching RDS, ElastiCache and S3 is already encrypted. Attested instances run in subnets with no internet egress, and inference servers never receive keys. Service Control Policies across Bee’s accounts add separation of duties: deploying unlogged code would need several independent teams to cooperate.
What This Means for You
The useful lesson is the split Bee draws. Attestation proves what runs inside an instance but can’t measure what surrounds it, such as network paths or who can change account permissions. So every control that attestation can’t see gets its own layer. AWS’s guide to isolating data from your own operators is the starting point, and Trail of Bits independently assessed Bee’s whitepaper.
The trade-offs are operational, in our read: with no shell there’s no logging in to debug a live instance, and every deployment depends on another team’s transparency service. Bee is moving inference to Amazon Bedrock Mantle, which offers zero operator access as a managed capability. For containers, AWS points to Nitro Sandboxes, which give each pod or task a hardware-isolated VM with attestation built in.
The So What
Platform teams that promise customers their staff can’t read sensitive data should test the claim. Can a client verify it, or does it rest on IAM policy? If it rests on policy, start with attestation-gated key release for the single service that handles plaintext. Hand the reference measurements to a team that can’t deploy, and remove shell access from those images first.
Source: AWS Architecture Blog
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →