Playbook · for Engineering, Security
Secure Amazon Bedrock in production
IAM, guardrails, invocation logging, private networking and encryption for Bedrock workloads, in the order that cuts the most risk.
By Mr.DLast reviewed 3 min read
Why Bedrock needs its own checklist
Amazon Bedrock removes the hardest part of running models, but not the ordinary cloud risks around them. The incidents we track rarely break the model. They use a role that can call any model, a leaked key, a public path to an endpoint, or the absence of logs that would have shown the abuse. Work through these steps in order: each one closes more risk than the next.
1. Least-privilege IAM for model access
Never attach AmazonBedrockFullAccess to application roles or people who only call models. It also lets them change guardrails, logging and custom models. Grant only what the application uses:
json{ "Version": "2012-10-17", "Statement": [ { "Effect": "Allow", "Action": ["bedrock:InvokeModel", "bedrock:InvokeModelWithResponseStream"], "Resource": "arn:aws:bedrock:eu-central-1::foundation-model/<model-id>" } ] }
- Scope
Resourceto the model ids and Regions you actually use. - Use a separate role per application, so one compromised service cannot run up costs for all.
- Prefer roles over long-term API keys. If you must use Bedrock API keys, keep them short-lived and out of code.
2. Turn on model invocation logging
Invocation logging records who called which model, when and with what input and output. It is how you notice LLMjacking, prompt-injection attempts and runaway agents, and it is the evidence an auditor asks for.
- Enable it per Region, to CloudWatch Logs, S3 or both.
- Encrypt the log destination with a KMS key you control, because prompts and outputs often hold personal or confidential data.
- Set retention deliberately and restrict who can read the logs.
- Alarm on unusual invocation volume, new models and calls from unexpected Regions.
3. Keep traffic private
Use VPC interface endpoints (AWS PrivateLink) for bedrock-runtime and, if you use agents, bedrock-agent-runtime. Traffic then stays on the AWS network, and you can attach an endpoint policy and deny calls that do not come through the endpoint with the aws:SourceVpce condition key.
4. Use guardrails, and know their limits
Bedrock Guardrails filter prompt attacks, denied topics and sensitive information such as personal data. Turn on the prompt-attack filter at a high strength and the sensitive-information filter for the data classes you hold. Then treat guardrails as one layer: they reduce attacks and noise, but they do not make prompt injection safe. The prompt injection playbook covers the design controls that do.
5. Encrypt what you store
Use customer-managed KMS keys for saved prompts (Prompt Management), agent resources, knowledge bases, custom model artifacts and the invocation logs above. A key you control lets you revoke access and shows an auditor who can decrypt.
6. Watch the account, not only the model
Bedrock runs in an AWS account, and that account is the real perimeter:
- CloudTrail in every Region, so control-plane changes such as disabling logging are recorded.
- GuardDuty enabled, to flag credential misuse and unusual API calls.
- No administrative policies on application roles, and no privilege-escalation paths.
- Budgets and cost anomaly detection, because abuse of model access shows up first on the bill.
What WardBee checks
WardBee connects read-only to AWS and checks Bedrock settings, such as full-access policies, Bedrock API keys, the prompt-attack and sensitive-information filters on each guardrail, guardrails on agents, invocation logging and its encryption, VPC endpoints and customer-managed keys for prompts, next to IAM, CloudTrail and GuardDuty. Each result becomes ISO 27001 or SOC 2 evidence. Whether your application sends every call through a guardrail is not visible in the account, so review that by hand with step 4.
What WardBee checks for this playbook
- IAM role does not have AmazonBedrockFullAccess managed policy attached
bedrock_full_access_policy_attached - Amazon Bedrock API key does not have administrative privileges, privilege escalation paths, or full Bedrock service access
bedrock_api_key_no_administrative_privileges - Amazon Bedrock long-term API key has expired
bedrock_api_key_no_long_term_credentials - Amazon Bedrock guardrail has prompt attack filter strength set to HIGH
bedrock_guardrail_prompt_attack_filter_enabled - Amazon Bedrock guardrail blocks or masks sensitive information
bedrock_guardrail_sensitive_information_filter_enabled - Amazon Bedrock agent uses a guardrail to protect agent sessions
bedrock_agent_guardrail_enabled - Amazon Bedrock model invocation logging is enabled
bedrock_model_invocation_logging_enabled - Amazon Bedrock model invocation logs are encrypted in the S3 bucket and KMS-encrypted in the CloudWatch log group
bedrock_model_invocation_logs_encryption_enabled - VPC endpoints ensure private connectivity for all Bedrock APIs
bedrock_vpc_endpoints_configured - Amazon Bedrock prompt is encrypted at rest with a customer-managed KMS key
bedrock_prompt_encrypted_with_cmk
Runs on each AWS sync while a framework it maps to, such as ISO 27001, is enabled. Run it on your account
Related standards: ISO/IEC 27001 · ISO/IEC 42001 · OWASP Top 10 for LLM Applications


