Bee, Amazon’s personal AI wearable, is designed to listen to your day, summarise conversations, and surface insights—handling some of the most sensitive data a user can generate. The stakes are high: it processes information about health, finances, and relationships, and the risk of operator access (whether by Bee engineers, Amazon employees, or AWS infrastructure staff) is unacceptable for its users. Bee’s solution: an architecture where there is no mechanism for any operator to access customer data—what they call Zero Operator Access (ZOA).
The core of Bee’s approach is EC2 instance attestation, backed by the AWS Nitro System, with additional network, storage, and organisational controls to cover gaps attestation alone can’t measure. Here’s how the pieces fit together.
Building Attestable Infrastructure
At build time, Bee’s pipeline produces hardened, attestable Amazon Machine Images (AMIs) for EC2. A separate Amazon team runs a transparency service: every approved AMI, container, and model is logged for audit, with expected measurements published to Sigstore Rekor, a public transparency log. These reference measurements are based on the Unified Kernel Image (UKI) and Secure Boot configuration, specifically targeting PCR4, PCR7, and PCR12 in NitroTPM.
Each AMI is built without interactive access—no SSH, SSM, or shell—so operators have no way into a running instance. The root filesystem is read-only, and local storage is encrypted with a key that only exists in memory. Instances are replaced instead of rebooted to ensure no data persists beyond the instance’s life.
Runtime: Attestation-Gated Key Release
When a user’s Bee device connects, it only talks to an EC2 instance that can prove it is running the approved software. At boot, the instance generates a NitroTPM attestation document, signed by AWS Nitro, and obtains a TLS certificate from AWS Private CA embedding its attestation and transparency proofs. The Bee app verifies these before releasing any decryption keys.
If the software on the instance has changed, the app refuses to connect. Even when connected, the key released is short-lived and only stored in memory on app servers and Redis; inference servers never receive the key. Data in Amazon RDS, ElastiCache, and S3 is encrypted before leaving these boundaries, and AWS Lambda integrations only see plaintext if the user consents.




