Security

Your GPUs, your data, least privilege.

AWSBee is a control plane. Inference runs on instances in your account; AWSBee holds only what it needs to launch, monitor and route to them.

Data plane isolation

Model servers run on EC2 instances in your account. Weights are downloaded from the publisher directly to your instance.
The model port (8000) accepts traffic only from the AWSBee gateway's fixed IP address, plus an optional CIDR you choose.
vLLM deployments require a per-deployment API key generated by AWSBee and passed to the stack as a NoEcho parameter.
Instances enforce IMDSv2 and use encrypted gp3 volumes. Stacks delete every resource they created.

The gateway streams requests and responses through without storing prompts or completions.
It records metadata only: timestamp, deployment, status code, latency, time to first token and token counts.
All public traffic uses TLS. Gateway keys are stored as SHA-256 hashes and shown once at creation.

Cloud secret keys, Hugging Face tokens and NGC keys are encrypted at rest with AES-256-GCM.
The IAM policy we publish is scoped to stacks named bee-*, the EC2 actions a GPU instance needs, read access to the public Deep Learning AMI parameters and Service Quotas reads. It cannot read your data, S3 buckets or other workloads.
Disconnecting your cloud account deletes the stored key immediately; we recommend also deactivating it in IAM.

Passwords are hashed with scrypt and a per-user salt. Sessions are random tokens; only their hashes are stored.
Sign-in and sign-up are rate limited per IP address.

Email security@awsbee.com. We acknowledge reports within two business days and do not take action against good-faith research.