What turns a policy into a control?
A policy is text. A runtime control is software that reads each prompt, response or tool call, decides whether it breaks the policy and acts: allow, block, rewrite, mask or escalate. The decision is logged, and the log becomes evidence.
Getting there needs three translations: from the policy's words to categories a detector can check; from categories to thresholds and actions; and from actions to a record. A platform that is 'built for your policies' lets you do those translations for your own application rather than using one generic list.
What do the six platforms describe?
- Alice: WonderFence sits between external-facing AI and its users and intercepts harmful, non-compliant and off-policy responses, configured to the application's policies, with every decision logged.
- Pillar Security: runtime guardrails block prompt injection, jailbreaking and tool manipulation, mask PII, PHI and secrets and log every prompt, response and tool call.
- Lasso: policy enforcement and detection and response, plus the CPU-based LEAP guardrail announced on 3 September 2026.
- SPLX: input and output guardrails, and system prompt hardening through Dynamic Remediation.
- Holistic AI: describes enforcing policies; the runtime mechanism is not described on the page we read.
- Credo AI: describes enforcing policy; runtime blocking is not described on the pages we read.
What about latency?
A control in the request path adds time. Vendors publish their own figures: the WonderFence page title states sub-150 ms, and Lasso says LEAP decides in under five milliseconds. These are vendor claims measured in the vendors' own conditions. Measure in yours during a proof of concept, with your traffic, your languages and your policy set.
What should you check?
- Can you write or tune policies for your own application, and in which languages and modalities?
- Which points does the control cover: input, output, tool calls?
- What actions are available besides block: rewrite, mask, escalate to a human?
- Is every decision logged with the policy version that made it?
- How do red teaming results change the control?
What happens after a control blocks something?
A block is the start of a record, not the end. The user needs a response that still follows the policy, such as a refusal or a handoff to a person. The log needs the input, the decision, the policy and its version. Someone needs to review a sample of blocks and allows for false positives and misses, and feed what they find into the next round of tests. Without that review, a control can drift away from the policy it was written to enforce.