Keeping evidence for Amazon Bedrock and Google Vertex AI calls when provider-side logging is off
When Amazon Bedrock model invocation logging and Vertex AI request-response logging are off, the provider holds no copy of your prompts and responses, so you must produce the evidence yourself. Define a per-call record (identity, timestamp, model and endpoint, policy decision, action taken, and redacted or hashed content), capture it at a point you control, and store it with your own retention, residency, integrity and access controls. Keep the native cloud audit trail as a baseline that each call occurred.
Define the evidence record before choosing a capture point
- Identity: the user, service or agent credential that made the call, plus the application or team it belongs to.
- Timestamp: a trusted time for the request and for the response, so ordering can be reconstructed.
- Model and endpoint: the model identifier, provider, region and endpoint actually invoked.
- Policy decision: which policy version applied and whether the result was allow, redact or block.
- Action taken: what happened to the call as a result, including any modification or refusal.
- Content treatment: the prompt and response stored redacted, hashed or excluded, with a record of which treatment was applied.
Where the evidence gap sits
Cloud audit trails such as AWS CloudTrail and Google Cloud Audit Logs record API activity: who did what, when and from where. That baseline matters. Without it there is no authoritative account that a call happened, and OWASP guidance recommends enabling audit logging in every region and account and verifying coverage of management events, data access events and security-relevant API calls.
Provider-side content logging is a separate matter. Regulated teams often turn off Bedrock model invocation logging and Vertex AI request-response logging to avoid storing prompts and responses in provider-managed destinations, or to meet data minimization and residency requirements. That is a reasonable decision, but it removes the provider-held copy of what was sent and what came back. An audit trail that shows a call occurred does not, on its own, show what the call was allowed to do, which policy applied, or what the model returned.
A NIST-hosted AWS incident-response presentation makes a related point: customers often lack the AI-specific evidence trail, including prompt and completion records. The gap is therefore a design problem for the customer, not a configuration fix. The evidence has to be produced by something you operate.
Regulation points the same way. The EU AI Act says high-risk AI systems shall technically allow for automatic recording of events (logs) over the lifetime of the system, and that providers keep those logs to the extent they are under their control. As of November 2026, whether a given system is high-risk, and which role your organization holds, determines how that applies. Still, evidence you generate and store yourself is plainly within your control. Provider logs you have disabled are not.
Handling content when raw prompts cannot be stored
If storing raw prompts is restricted, evidence does not have to mean storing them. Three treatments cover most cases. Exclusion records the metadata and decision but no content. Redaction stores the content with sensitive values removed or replaced, and notes what was removed. Hashing stores a digest of the content, which lets you later show that a specific prompt or response matches the record without keeping the text.
Choose per data class, not globally. A hash proves a match only if someone can present the original, so it supports verification rather than review. Redacted content supports review but depends on the quality of the redaction. Record the treatment in each entry so a reviewer knows what the evidence can and cannot show.
The most defensible record is one written at the moment the decision is made. Reconstructing a decision from application logs afterwards leaves room for disagreement about what actually ran.
Comparing capture points
There are three common places to produce the record. This comparison is editorial analysis, and the right choice depends on how many teams call the models and how much bypass risk you accept.
- Application-layer instrumentation: Each application or SDK wrapper writes the record. Coverage depends on every team adopting it correctly, and any code path or credential that skips the wrapper leaves no record. Latency impact is small, and retention is under your control. Policy decisions are only recorded if the application makes and reports them.
- Proxy or gateway pattern: Calls route through a shared intermediary that records them. Coverage is broader and consistent, but only if direct access to the model endpoint is closed off. It adds a network hop and a component to operate. Whether it records a policy decision, or only traffic, depends on what it is configured to do.
- Enforcement in the execution path: Policy is enforced inline on each interaction, and the evidence record is a by-product of that decision. A call cannot proceed without passing the decision point, so the record and the action are tied together. It needs the same bypass controls as a proxy and adds inline latency to every call.
Implementation sequence
Work through these steps in order. Each one assumes the evidence record above is already defined, and each applies equally to Bedrock and Vertex AI.
- 1. Secure the baseline: Enable cloud audit logging in every region and account, covering management and data access events, so a record that each call occurred exists regardless of content logging.
- 2. Close bypass paths: Restrict permission to invoke the model endpoints to the identity used by the capture point. Check that no team or credential can reach Bedrock or Vertex AI another way, and that capture cannot be silently disabled or skipped.
- 3. Set content handling: Assign exclusion, redaction or hashing to each data class and write the treatment into the record.
- 4. Place storage and set retention: Store evidence in a region that satisfies your residency requirements. Set the retention period from the regime that applies to you, and plan the full lifecycle of generation, transmission, storage, access and disposal.
- 5. Protect integrity: Evidence can be deleted by an attacker hiding their activity. Keep the store in a separate account from the workloads it observes, restrict deletion, and use write-once storage or signed or chained entries so tampering can be detected.
- 6. Control access and assign responsibility: Limit and review who can read evidence, since a competent authority may request it. Document which party, provider or customer, produces and retains each item of call evidence.
Tradeoffs and where Trussed fits
No capture point is free. Application instrumentation is simple to start but hard to prove complete. A proxy gives consistent coverage but becomes a dependency and a target. Inline enforcement gives the strongest link between decision and record, at the cost of sitting in the latency path of every call. In each case, the evidence is only as good as your ability to show that calls cannot go around it.
Trussed enforces policy in the execution path, with allow, redact and block decisions applied at runtime, not reviewed after the fact. That enforcement produces audit-ready evidence, mapped to HIPAA, NAIC, NYDFS, the EU AI Act and NIST. It can be self-managed in your own cloud, with configurable retention and data residency, which suits teams that disabled provider-side logging to keep content in their own boundary. Trussed scopes this to individual interactions and tool calls.
Whichever approach you choose, test it the way an auditor would: pick a call, and show who made it, which policy applied, what the decision was and what happened next.
Evidence path for one model call
Caller
Identity and credential that made the request.
Capture point
Where the policy decision is made and the record is written.
Model endpoint
Bedrock or Vertex AI, called only through the capture point.
Evidence store
Customer-controlled retention, residency, integrity and access.
Show what each AI call was allowed to do
See how policy enforced in the execution path produces audit-ready evidence in your own cloud, with the retention and residency you set.
See How Evidence Is Produced