Restricting an EDD Agent's Adverse Media Searches to Approved Sources
To restrict an Enhanced Due Diligence agent to approved sources, define the allowlist outside the prompt, remove every tool that can reach the open web without passing through it, and enforce the allowlist on each search and retrieval call at runtime. Out-of-scope requests should be blocked or escalated to a human approver, and every decision should produce a record of the policy version, the request and the outcome.
Step 2: Map the list to the agent's search and retrieval path
- 1
Search API configuration
Restricting the provider to named sites is simple, but it covers only that one tool and depends on provider features.
- 2
Retrieval-layer filtering
Filters what is indexed or returned, but does not stop other tools from fetching directly.
- 3
Network egress controls
Blocks unlisted hosts broadly, but cannot see intent, tool identity or policy version.
- 4
Tool-call policy check
Evaluates each call against the allowlist before execution, across every tool that is routed through it.
Step 4: Capture audit-ready evidence for each search
- Policy version and effective date in force for the call
- Tool invoked, requested destination and final resolved destination
- Decision taken: allow, redact, block, rate limit or escalate
- Approver identity, time and rationale for any escalation
- Timestamp and the case or review the request was made for
- Retention period and storage location for the record
Why prompt instructions do not hold the boundary
An instruction such as "only use approved sources" tells the model what is expected. It does not stop a tool from executing. OWASP Cornucopia lists three conditions that lead to agent scope violations: missing context boundaries, weak prompt enforcement where system instructions are not structurally protected, and over-permissive tool access. It also notes that when an agent cannot tell trusted instructions from untrusted content, injected content can push it outside its intended task. Adverse media search is exposed to this by design, because the agent reads text from third-party pages.
NIST's generative AI profile makes a related point about output restriction: it may still fail against novel prompts. A draft NIST publication on adversarial machine learning describes agent mitigations as constraining and monitoring the agent's actions against policies set by the deployer. That is the framing used here. Source restriction is a property of what the agent is allowed to call, enforced in the execution path, not a property of what it was told.
The policy itself is also an institution's design choice. FinCEN has stated, citing earlier guidance, that compliance with the CDD Rule does not categorically require media searches or particular screenings. The approved-source list should therefore be framed as your risk-based method, owned by compliance and written down, rather than as a universal requirement.
Step 1: Build the approved-source policy
Start with an explicit allowlist, consistent with the principle of integrating only vetted and authorized components. Compliance decides which sources qualify and why. The reliability criteria are theirs to set. Security's job is to turn the decision into something a machine can evaluate.
Treat the list as a versioned artifact, so that any past search can be tied to the exact policy in force at the time.
- Source identifiers: Record each approved source as a domain, URL pattern or named API endpoint, not a description.
- Approval metadata: Store who approved the source, when, and the stated rationale.
- Scope conditions: Note any limits, such as permitted query types or jurisdictions, that the policy can check.
- Version and effective date: Every change creates a new version; old versions stay readable.
- Default rule: State that anything not listed is denied or escalated, never silently allowed.
Step 3: Write allow, block and escalate rules
Enforcement at the tool call lets you decide per request. A workable starting design has three outcomes. Allow requests whose resolved destination matches an approved source and whose parameters are within scope. Block requests to destinations that are clearly unapproved, or from tools that should not be used for this task. Escalate requests that are ambiguous, such as an unlisted source the agent argues is relevant, so a named approver can decide.
This three-way split is a design choice, not a published standard. Its value is that the agent never has to decide for itself whether a source is acceptable.
Two supporting controls are worth adding. Rate and volume limits on queries are a recognised usage restriction in NIST draft guidance, and they cap the damage from a looping or manipulated agent. Where policy permits, redaction can remove identifying details from outbound queries before they leave your environment.
Escalation needs its own rules. Define who may approve, how long a request waits, and what the agent does while waiting. Decide whether an approval applies to one call or adds the source to the list through the normal compliance review. One-off approvals should not quietly become permanent exceptions.
Testing the boundary and where Trussed fits
Test the control as you would any access control. OWASP recommends penetration testing and breach simulation that treat the model as an untrusted user. For this use case, plant instructions in a retrieved page that tell the agent to search elsewhere, then confirm the call is blocked or escalated. Test redirects, aggregator pages and chained tool calls. Keep monitoring continuously afterward, since a federal RFI on agent security lists tool restrictions and continuous monitoring as controls while openly asking how mature they are. Avoid treating any single method as settled.
Trussed enforces policy in the execution path on individual interactions and tool calls, with allow, redact, block, rate limit and escalate for human approval outcomes. It generates audit-ready evidence from that enforcement. That maps to steps 3 and 4 above. It applies per call, so it does not by itself provide end-to-end tracing of a whole agentic workflow. The approved-source list and the reliability criteria remain decisions for your compliance team.
Approved-source enforcement path
Define
A versioned allowlist of approved sources, owned by compliance.
Map
Every agent tool that can search or retrieve, reduced to what the task needs.
Enforce
Allow, block or escalate each call before it executes.
Evidence
A record of permitted scope and actual requests per search.
Enforce source scope where the agent acts
See how policy enforcement on individual tool calls and the audit-ready evidence it produces apply to an EDD agent's search path.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.