self-hosted-llm-batch-inference
Pass
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: SAFE
Full Analysis
- Secure Credential Management: The skill emphasizes the use of AWS Secrets Manager for staging model credentials and explicitly warns against passing long-lived tokens to the runtime environment, adhering to the principle of least privilege.
- Data Integrity and Verification: It prescribes mandatory checksum verification for model artifacts, input manifests, and individual shards (SHA256) to ensure that the data being processed has not been tampered with or corrupted during transfer.
- IAM Boundary Enforcement: The guidance explicitly defines separate IAM roles for staging, job execution, and infrastructure management. This modular approach ensures that a compromise in the inference container does not grant access to modify infrastructure or model weights.
- Controlled Input Processing: By requiring deterministic input shards and a strictly defined output/quarantine schema, the skill reduces the surface area for unexpected behavior during inference and provides a clear path for handling malformed data without failing the entire run.
- Secure Infrastructure Configuration: It utilizes AWS native services (AWS Batch, ECS Managed Instances) and provides specific configurations, such as setting
ScaleInAfterto zero and removing inbound security group rules, which minimize both costs and the attack surface.
Audit Metadata