Prompt VaultExamples

Prompt Vault showcase

Prompt Vault gives enterprise customers a concrete answer to the raw-prompt visibility question: AgentKeeper can detect risky prompts without making every prompt visible in normal telemetry.

Prompt Vault gives enterprise customers a concrete answer to the raw-prompt visibility question: AgentKeeper can detect risky prompts without making every prompt visible in normal telemetry.

The customer story

Companies often want AI runtime governance, but their prompts can contain product strategy, customer data, legal questions, source code, credentials, or incident details. Prompt Vault separates enforcement from retention:

  1. AgentKeeper receives a supported prompt event from a runtime source.
  2. Runtime Shield evaluates the full prompt in memory.
  3. The organization's Prompt Vault mode determines what is retained.
  4. Activity, findings, investigations, exports, and OTLP APIs render the governed form, not an accidental raw copy.

Three modes

ModeBest forUser experience
Visible CaptureInternal pilots, security labs, high-context investigationsPrompt text appears normally to authorized users.
Sealed EvidenceEnterprise production environments that still need break-glass reviewPrompt text is hidden by default. Admins see metadata and a Reveal button.
PromptlessHighly restricted environments where raw prompts must not be retainedPrompt text is evaluated in memory and discarded. No reveal is possible.

What admins see

Settings has a first-class Privacy tab with Prompt Vault controls. Admins choose a mode, read the operational impact, and review reveal logs from the same place.

In evidence views:

  • Visible Capture shows the prompt body like existing Activity evidence.
  • Sealed Evidence shows source, length, hash, timestamp, and a governed Reveal prompt button.
  • Promptless shows that the prompt was evaluated but not retained.
  • AgentKeeper Runtime sensor rows, including eBPF-style process and connection observations delivered through OTLP transport, show as runtime evidence. They do not show Prompt Seal controls unless an actual prompt evidence record is linked.

The same state appears across Activity drawers, Workstation event cards, Prompts & Decisions, Findings, and Investigation evidence so analysts do not have to learn separate meanings per page.

Break-glass reveal

Sealed Evidence is the main enterprise mode. Revealing a prompt requires:

  • owner or admin role,
  • a reason entered before access,
  • an access log row written before the encrypted prompt is returned.

The reveal log records who accessed the prompt, when, from where, and why. It does not copy the prompt body into the log.

Customer answer

When a customer asks, "Can RAD see our raw prompts?", the concrete answer is:

  • Visible Capture: yes, authorized dashboard users can see retained prompts.
  • Sealed Evidence: not in normal telemetry. A privileged reveal with reason is required and audited.
  • Promptless: no, raw prompts are not retained.

Related detail: What AgentKeeper sees, admin guide, and technical coverage.