Architecture-specific review

Amazon Bedrock security review

Bedrock supplies model access and security capabilities. A production review still has to establish the application's authority chain, data paths, agent actions and evidence for each required control.

Request access →

Review scope

What Drel establishes for this stack

Review the workload that calls Bedrock, its AWS accounts and connected systems. A configured guardrail or VPC endpoint addresses a specific boundary; neither should be treated as proof that every tool call or retrieved document is safe.

Drel rebuilds the system from its repository, maps the attack paths through each area, checks the controls it can verify in your code, and proposes the fix for each control that fails. A reviewer decides what counts as evidence.

01

IAM and model invocation

Which principals can invoke which models and change guardrail or logging configuration? Are application actions constrained independently of model access?

What counts as evidence: IAM policies and roles, trust relationships, model resource scope, change permissions and the application authorization path.

02

Network and data movement

What network path does the workload use, and where do prompts, outputs, retrieved content and invocation logs go?

What counts as evidence: VPC endpoint and endpoint policy where used, data-flow diagram, storage access, encryption choices and logging configuration.

03

Retrieval and untrusted content

Can retrieved or user-supplied content redirect the agent or expose another user's data?

What counts as evidence: Ingestion boundaries, retrieval filters, source attribution, prompt construction, and the attack paths Drel maps from the design to the data.

04

Guardrails and tool permissions

Where are guardrails applied and enforced? Can the agent take actions outside the user's authorized scope or bypass a required approval?

What counts as evidence: Guardrail attachment and IAM enforcement, tool schemas, server-side permission checks, human approval and audit events.

05

Observability and approval

Which events are logged without leaking sensitive payloads? What gaps remain and who accepts the conditions of release?

What counts as evidence: Invocation and audit logging settings, access to logs, test results, control owners and the signed review decision.

Architecture example

Follow authority and data end to end

A model endpoint is one component. The security decision also depends on the identity that calls it, the data it receives, the actions it can trigger, and who accepts the remaining risk.

For a claims assistant: employee request → application identity → Bedrock model invocation → retrieval from an approved data source → proposed claim update. Trace both the employee's right to see the claim and the service role's right to modify it. A model instruction cannot grant either permission.

From repository to approval

Drel models the system from its repository, threat-models every attack path against OWASP, the EU AI Act and 22 more, maps the attack paths through it, checks the controls it can verify in your code, proposes the fix in a Jira ticket, and re-checks the automatic checks you turn on when the branch you link changes. Your AI Committee signs the approval. Unknown or missing evidence remains visible; it is not treated as a passed control.

Read the review methodology →

Further reading