Skip to content
be-human.ai
← Back to the journal

AI by industry

AI Agent for Cybersecurity Consulting: Separate Evidence Queues

Scope audit evidence queues and request logs per client so an AI agent never mixes one engagement's findings into another's. Plain rules for consultancies.

Be Human typographic cover with the title: AI Agent for Cybersecurity Consulting: Separate Evidence Queues
Conceptual illustration created with AI · Be Human

An AI agent for a cybersecurity consulting firm should work from one evidence queue per client engagement, with its own memory namespace, retrieval index, tool permissions and request log. Scan exports, screenshots, interview notes and draft findings stay inside the boundary of the engagement that produced them, even when the same agent supports several clients. Scoping these queues before the agent becomes more capable is what keeps one client's findings out of another client's report.

Key takeaways

  • Give every client engagement its own evidence queue, memory namespace and request log.
  • Treat cross-client retrieval as an authorization decision, not a search convenience.
  • Log each request with the engagement, the requesting reviewer, the tool used and the artifacts returned.
  • Keep draft findings separate from released report text until a named reviewer approves the release.
  • Test isolation with planted canary records before any real client material enters the system.

Why evidence queues matter more as agents get stronger

A consultancy's most sensitive material is rarely its methodology. It is the evidence gathered during each engagement: configuration snapshots, scan outputs, interview summaries, screenshots of internal systems and findings that have not been checked yet. A capable agent is useful because it can connect these items quickly. That same ability is the risk. An agent that can search everything it has access to will eventually surface a fact from one client while answering another.

Prompt instructions such as 'never mention other clients' are requests to the model, not controls. The material the model reads can contain text that tries to override those requests, which is why the OWASP guidance on prompt injection treats untrusted content as something to constrain by design. For a consultancy, the boundary between clients has to live in storage, retrieval and tool permissions, where it can be tested and audited. Firms that already plan memory across several client accounts, as described in our note on managed service provider memory, face the same structural question.

Define the engagement as the unit of isolation

Start with the engagement, not the client. One client may have an annual penetration test, a retest after remediation and an unrelated advisory project, and each of these needs its own boundary. Every engagement record should hold the statement of work, the rules of engagement, the authorized asset list, the evidence queue, the request log, the named reviewers and a retention and deletion date. The scoping work itself follows a sequence that can be checked item by item:

  • List every active, closed and retention-pending engagement, and give each an identifier that does not reuse the client name.
  • Write the authorized scope as structured records: assets, address ranges, environments and testing windows.
  • Create one evidence queue per engagement, with a separate storage path and a separate retrieval index.
  • Assign each memory namespace to a single engagement, and keep shared playbooks in their own namespace, limited to material cleared for reuse.
  • Map which tools the agent may call in each engagement, and mark which of them can read, write or release content.
  • Set retention and deletion dates at intake so closed engagements leave working memory on schedule.
  • Run a canary test: plant a unique marker in one engagement and confirm that no other engagement can retrieve, quote or log it.

The canary test is the step most often skipped. It turns the isolation claim into a result that someone can reproduce after the next change to the retrieval setup.

Illustrative scenario: the same weakness at two clients

Illustrative scenario: a consultancy runs two external assessments in the same month. Both clients operate an older remote access appliance from the same vendor, and the agent, reviewing scan outputs for each, notices the pattern in both. A well-scoped agent can state that the appliance family has public advisories and that the reviewer should check both assets against them, citing the public advisory. It should not say that one client's appliance showed a particular configuration, or that the other client's engineer accepted a risk, because those facts sit in a different engagement's queue.

If the reviewer asks what the other engagement found, the correct response is that the agent has no access to that material in the current context. The attempt should then appear in the request log as a denied cross-engagement lookup. A refusal that leaves no trace tells the consultancy nothing about how often the question came up, and that count is often the first sign that a scope or a reviewer's workflow needs adjusting. The same discipline appears in other field-based services, such as the post orders and incident logs described for security guard firms.

Build the request log so it can answer who saw what

The request log is the record that clients, auditors and the consultancy's own quality function will eventually ask for. Observability practice treats a single operation as a trace made of connected steps, and applying that idea to agent actions gives you a path from each request to the evidence it returned. The OpenTelemetry concepts guide is a useful reference for the vocabulary. For each agent action, record:

  • The engagement identifier and the requesting reviewer.
  • The tool or retrieval index used, with its version.
  • The identifiers of returned artifacts, never the artifact text.
  • Denied requests and the rule that blocked them.
  • Release events, with the person who approved the release.

Store the log where the agent cannot edit it, and keep it append-only. Review it weekly during active engagements and again at closure. Repeated denials often mean a reviewer needs a different scope, and that is better fixed in the statement of work than worked around in a prompt. Quality teams already treat audit trails this way, as in the audit preparation work that depends on traceable nonconformance records.

Permissions, drafts and release to clients

Separate three states clearly: draft inside the engagement queue, reviewed internally and released to the client. The agent can create drafts within its own engagement. Release should always require a named human, and that step should write its own event to the log. Permission checks should run on every call, deny by default, and take the engagement identity from the authenticated session rather than from anything the agent says about itself. The OWASP authorization guidance describes this pattern in general terms, and the NIST AI Risk Management Framework is a useful reference for assigning clear roles to the people who govern each step.

A practical permission matrix for a consultancy often looks like this:

  • Read: the current engagement's queue and playbooks cleared for reuse.
  • Write: drafts in the current engagement only.
  • Release: not granted to the agent. A reviewer performs every release.
  • Delete: limited to retention workflows, with a second approver.

When a scope changes mid-engagement, the permission record changes with it, and the change is logged with the reason.

How Be Human's managed approach handles this

Be Human designs the engagement boundary during the first engagement, defining the memory namespaces, tool permissions and approval steps, then implements the queues, indexes and logs so the boundary can be tested. In managed operations, the work can include monitoring of request logs, recurring canary tests for isolation, permission reviews whenever an engagement scope changes, incident response if a boundary fails, and controlled improvements to retrieval that must pass the isolation tests before they reach client work. The service is remote, in English or Portuguese, and the configuration follows the consultancy's own rules of engagement rather than generic defaults. Be Human takes the free AI diagnostic to find the first task in your consultancy worth handing to an agent.

Frequently asked questions

Can one AI agent serve several consulting clients without mixing their findings? Yes, but only when each engagement has its own evidence queue, memory namespace and retrieval index, and permissions are checked against the engagement on every call. Prompt instructions alone are not a substitute for that structure. Shared material should be limited to public sources and playbooks that have been cleared for reuse.

Does a more capable agent increase isolation risk? It can, because a stronger model connects information across more sources and handles more context at once. Isolation should therefore be enforced in storage, retrieval and tool access, and tested with planted canary records each time the capability changes. Every upgrade should rerun the same isolation tests before it touches client work.

What should a request log record for a cybersecurity engagement? Record the engagement identifier, the requesting reviewer, the tool or index used, the identifiers of returned artifacts, denied requests with the rule that blocked them, and any release event with the approving person. Keep the artifact text out of the log itself, since the log will be read by people who need to see patterns, not contents. Make the log append-only so it can answer questions after an engagement closes.

When should a consultancy bring in an outside operator? Consider it when one agent touches several engagements, when reviewers need different permission levels, or when retention and deletion obligations apply to client material. An operator can run monitoring, isolation tests and permission reviews on a recurring schedule. Decisions about what a client may see remain with the consultancy's own reviewers.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →