TL;DR: Almost every vendor claims runtime enforcement, so compare AI governance tools and runtime guardrail platforms on depth: where enforcement sits, whether it blocks or only detects, and whether the feature is generally available. This guide compares seven vendors on enforcement, deployment, agent identity, kill switch and spend limits, and SIEM export. Prediction Guard, a sovereign AI control plane, is one of the seven, and its limits are stated.
Key takeaways
- Almost every vendor can claim "runtime enforcement." The useful questions are where the enforcement point sits, whether it blocks or only detects, and whether the feature is generally available or still a preview.
- Enforcement depth runs on a ladder: pattern filtering, semantic detection, tool-call validation, and full agent-chain enforcement. Most products concentrate on the first two rungs.
- Only a few vendors document agent identity, a way to stop an agent, and spending controls together, and the controls differ in scope. Some enforcement features remain in preview or research preview.
- Where the product runs, and how it is integrated, decides where your keys, prompts, and evidence live, which is often what an examiner asks about first.
Regulated teams comparing runtime enforcement products and AI governance tools meet the same problem: vendors describe very different things with the same words. This guide compares seven vendors on five questions: whether runtime enforcement blocks inline or only detects, where the product is deployed, what it does for agent identity, whether it has a kill switch and spend limits, and how audit records reach a SIEM. Prediction Guard, a sovereign AI control plane that runs inside your boundary, is one of the seven.
The seven come from different product categories, including a screening API, AI security products, governance platforms, and agent control planes. They are compared together because regulated buyers evaluate all of them for the same outcome, controlling what agents do at runtime. The central goal is operational control of agents: knowing what each one may do, being able to stop it, and being able to prove both.
A disclosure: Prediction Guard publishes this guide and builds one of the seven, so it is held to the same questions and its limits are stated. Every statement about another vendor comes from that vendor's own public pages as read in November 2026. A cell that says "Not publicly documented" means the vendor's public pages do not establish the capability, not that it does not exist. Nothing here is legal advice.
Why a Runtime Enforcement Claim Is Not Enough
Vendors market runtime enforcement as AI guardrails, and the phrase now covers everything from a keyword filter to a rules engine that decides whether an agent's next tool call may run. What separates these products for a regulated enterprise is depth: how far into a request, and how far into an agent's chain of actions, enforcement reaches before it stops. A responsible AI program is only as strong as the controls behind it, and the test of those controls is provable: a blocked call, a stopped spend, and a revoked identity, shown in the same session.
The Depth Ladder
Runtime enforcement splits into four levels, and many products concentrate on the first two.
- Pattern filtering. String matching against a denylist. Easy to bypass with a rephrase, and most modern products no longer rely on it alone.
- Semantic detection. A trained classifier evaluates intent and meaning, catching a rephrased injection or a PII entity described in natural language. This is where most AI safety enforcement products sit today.
- Tool-call validation. The product evaluates the specific action an agent is about to take: which tool, which parameters, which data scope, regardless of the reasoning behind it.
- Agent-chain enforcement. The product keeps enforcement context across multiple steps and tool calls, including calls routed through MCP servers, and catches a chain of individually plausible actions that adds up to something the agent was never authorized to do.
Where the enforcement point sits matters as much as the rung. A screening API sees the content it is sent. A gateway sees the traffic that passes through it. A hook inside an agent's harness sees each action before it runs. They answer different questions even when all three are called runtime enforcement.
Self-Hosted, Own Cloud, and SaaS: What the Labels Mean
Vendors use these three labels loosely, so this guide fixes them by who operates the software:
- SaaS: the vendor operates the service. What reaches the vendor depends on how the product is integrated and what it is configured to inspect or retain, so SaaS does not by itself mean that every prompt, tool call, and audit record goes to the vendor.
- Own cloud: the vendor's software runs in a cloud account or virtual network that you own, with the vendor operating or helping to operate it, often called hybrid or bring-your-own-cloud.
- Self-hosted: you operate the product yourself on infrastructure you control, such as your own data center or private cloud. Self-hosted does not by itself mean air-gapped, which is a further case with no internet connection and counts only where a vendor documents it.
The table reports each vendor's documented deployment, not a data-flow guarantee. In the pages reviewed, none of the seven documents a vendor-operated deployment in a customer's own cloud account, and the "private cloud" in Lakera's and Credo AI's pages is a place you self-host, so those cells say self-hosted. Of the seven, only Lakera's pages document air-gapped deployment.
The Comparison: AI Governance Tools and Runtime Enforcement Products Side by Side
Rows are alphabetical. "Documented" means the vendor's own public pages describe it, and each row's sources are listed below the table.
Sources for the Table
Credo AI: Agent Governor page, AI Agent Registry page, Credo self-hosting docs. IBM: agent control plane page, Agent Identity announcement, October 2026 announcement, agentic control plane announcement, governance enforcement announcement, Orchestrate deployment docs, watsonx on AWS announcement. Lakera: Agent Behavior Defense docs, integration guide, Agent Discovery docs, Lakera self-hosting docs, SaaS policy docs. Lasso: AI-DR page, MCP security page, AI gateway security page, agent governance page, MCP gateway repository. Microsoft: agent actions page, Foundry availability docs, service description, Agent 365 overview, Copilot Credits spending docs. Prediction Guard: its Platform v1.2.0 and Agents v1.2.0 release notes, its Admin Console documentation for API key quotas and telemetry export, and its blast radius guide for the documented limits. Prompt Security: AWS Marketplace listing, agentic security page, kill switch blog, acquisition announcement.
What the Table Shows
Enforcement points differ. Credo AI enforces inside the agent's own harness, Lasso at an API, firewall, or AI gateway, Prompt Security at an MCP gateway, browser, or plugin layer, Lakera screens content an application sends it, Microsoft and IBM through their identity and cloud layers, and Prediction Guard through the API gateway and agents of a platform you host. On the ladder, harness and gateway approaches reach tool-call validation most directly, while a screening API starts at semantic detection and adds tool-level checks through configuration.
Detecting and blocking are different jobs. Lakera's Tool Allow/Deny List returns a detection that your application must act on, while Credo AI, IBM, Lasso, Prompt Security, and Prediction Guard describe the product itself blocking or escalating. Microsoft's Defender blocks in preview.
Status changes fast. Credo AI's Agent Governor is a research preview and Microsoft's real-time protection is in preview, while IBM described a kill switch as tech preview on 27 October 2026 and listed one as new in its Agent Identity general availability release for watsonx Orchestrate on 3 November 2026.
Stop mechanisms and spend limits are the least uniform. IBM's kill switch undeploys an agent and ends its active sessions, Microsoft's Block action is organization-wide and its reach varies by agent type, and Prediction Guard's quarantine halts an agent's work. On spend, IBM documents cost tracking and cost control but no clearly established hard spending limit. Microsoft's Copilot Credits spending policies set limits for supported users, environments, and agents, but its pages describe them for Copilot services and do not present them as an Agent 365 feature. Prediction Guard enforces token and request quotas on each API key and refuses over-quota requests, though per-agent call budgets are not yet generally available. Lakera, Lasso, Prompt Security, and Credo AI do not publicly document a spend control.
Hosting and SIEM export decide the evidence path. IBM Orchestrate, Lakera, Credo AI, and Prediction Guard document a self-hosted option, Prompt Security does for three of its product lines, Microsoft runs as a cloud service, and Lasso does not state its platform's hosting. Lakera and Lasso say detections or findings can reach a SIEM, IBM says raw event data can go to existing security tooling, and Prediction Guard exports audit events through an OpenTelemetry collector endpoint and log shipping. Where the keys and the evidence live decides what you can produce for an examiner.
Enforcement depth is the thread through all of it. Screening a prompt or response sits on the first two rungs of the ladder, validating a tool call on the third, and controlling an agent across multiple steps on the fourth, and each needs a different enforcement point. A platform that blocks unsafe content does not necessarily control what an agent can access or do next. Buyers should evaluate enforcement depth alongside identity, delegated authority, and containment.
How to Choose a Runtime Enforcement Platform
Start from what you need to control, then check which vendors in the table document it.
- If you mainly need prompt and response screening, evaluate products built around detection and content filtering. In this table that is Lakera's screening API, with Detect and Enforce modes, as SaaS or a self-hosted container.
- If you need enforcement across multiple agents and tools, prioritize tool-level enforcement and agent identity: Credo AI's Agent Governor inside coding-agent harnesses (a research preview), Lasso or Prompt Security at an MCP or AI gateway, or Microsoft Agent 365 if your agents live in Microsoft 365 and Entra, noting that its real-time protection is in preview.
- If you operate in a regulated environment, examine deployment, data residency, and audit evidence first. IBM, Lakera, Credo AI, and Prediction Guard document self-hosted deployment, and Lakera also documents air-gapped operation.
- If you need to control autonomous agent activity, look for inline blocking, delegated authority, a kill switch, and spend limits together, and check the spend column closely, because the controls behind that label differ the most.
For AI risk management and model compliance programs, the deciding evidence is the same in every case: a blocked call, a stopped spend, a revoked identity, and an exportable audit record. That is operational control of agents, and it does not depend on what the product calls itself. For a longer view of the criteria, see Prediction Guard's key factors guide and its governance tools guide.
Where Prediction Guard, a Sovereign AI Control Plane, Fits, and Where It Does Not
A sovereign AI control plane is the architecture that identifies, authorizes, enforces, meters, records, and contains agent actions, running inside your boundary. Prediction Guard is one. What it ships today:
- Identity: one identity system covers users, roles, API keys, and agents across the Console, the API gateway, Agent Control, and Agent Forge, and each agent is a managed identity with an owner. API keys can be scoped to agents, knowledge bases, and skills.
- Delegated authority: a run is limited to what the API key, the agent, and the person who started it can all reach, including agents that call other agents.
- Containment: admins can quarantine an agent, which halts its work and prevents changes, and releasing it requires a reason. Disabling a user in one step ends their sessions, stops their API keys, and cuts off their agents.
- Resource control: token and request quotas on each API key, enforced at the gateway, with over-quota requests refused.
- Safeguards: PII, prompt-injection, toxicity, credential, and grounding checks, with a dashboard that charts blocked versus detected events per detector. A Tool Content control extends PII, prompt-injection, and secrets checks to tool results.
- Tool approval: admins can set each Agent Forge workspace file tool to Allow, Ask, or Deny.
- Deployment: self-hosted and managed from the Console through Cluster Settings and in-console upgrades with pre-flight checks, with Prediction Guard SaaS also offered and a migration path that carries external-model configuration.
- Audit: quarantine and release actions are recorded, a removed agent's record is kept, the live audit log has its own System tab, and audit events can be exported to a SIEM through an OpenTelemetry collector endpoint and log shipping.
These matter together because each answers a different part of the same question. Identity decides who is acting, delegated authority limits what a run can reach, quotas cap what it can spend, safeguards decide whether a request is blocked or only detected, and quarantine stops an agent while the record is kept.
Its documented limits apply. It enforces on traffic routed through it. Per-invocation authorization at the tool server, per-agent call budgets, and human-in-the-loop approval gates with workflow automation are not yet generally available; approval today is the Ask setting on workspace file tools and the Sessions API. Grounding reports and does not block.
How Prediction Guard Solves This
Prediction Guard is the sovereign AI control plane in this comparison. It runs inside your boundary and gives each agent a managed identity with an owner, limits every run to what the key, the agent, and the initiating user can all reach, enforces token and request quotas on each API key, can quarantine an agent, and exports audit events to your SIEM.
It enforces on traffic routed through it. Per-invocation authorization at the tool server, per-agent call budgets, and human-in-the-loop approval gates are not yet generally available. Ask us to show a blocked call, a stopped spend, and a revoked agent identity in the same session.
FAQ: Runtime Enforcement Platforms Compared
1. What is a runtime AI guardrail platform?
It is the market term for a product that checks, and in some cases constrains, what an AI application or agent does while it is running: the prompts it receives, the responses it returns, and in some products the tool calls it makes. Products differ in where the check happens (a screening API, a gateway, a hook inside the agent, or a control plane) and in whether a flagged action is blocked or only logged.
2. What's the difference between runtime detection and runtime enforcement?
Detection identifies a problem and records or flags it. Enforcement acts on it by blocking, rewriting, or escalating before the action completes. Lakera's Detect mode logs without blocking while its Enforce mode returns a flag that the calling application acts on, Credo AI's Agent Governor resolves each event as block, allow, escalate, or advise, and Prediction Guard charts blocked versus detected events per detector, with grounding reporting and not blocking.
3. Can AI guardrails prevent an AI agent from executing unauthorized tool calls?
Some can, depending on where enforcement sits. Lakera's Tool Allow/Deny List detects calls to tools outside an allow list and returns the result for your application to act on, Lasso can block or quarantine an action before it executes, Prompt Security's MCP gateway enforces allow, block, filter, or redact, Credo AI's hooks run before every tool call in a research preview, and Microsoft's Defender can block tool invocations. Prediction Guard limits every run to what the API key, the agent, and the initiating user can all reach, lets admins set Agent Forge workspace file tools to Allow, Ask, or Deny, and extends its prompt-injection, PII, and secrets checks to tool results. A flag that the application ignores prevents nothing, so what stops the call depends on where the enforcement sits.
4. What is the difference between an AI gateway and an AI control plane?
An AI gateway routes and meters traffic to models and tools. A control plane decides what each agent may do, enforces it, can stop an agent, and records the evidence. The categories overlap, and Microsoft and IBM both describe their products as control planes, so compare capabilities and not names.
5. Can runtime AI guardrails be deployed in a private cloud or air-gapped environment?
Some can, and the two are separate: self-hosting means you operate the product on infrastructure you control, while air-gapped means it runs with no internet connection. Credo AI documents self-hosting in your private cloud on Kubernetes or VMs, Lakera documents self-hosting in a data center or private cloud including air-gapped, IBM documents client-managed on-premises deployment, and Prediction Guard is self-hosted. SentinelOne's marketplace listing offers self-hosted Prompt Security for three of its product lines, Lasso's own pages do not state its platform's hosting, and Microsoft Agent 365 is a cloud service. Of the seven, Lakera is the one whose pages document air-gapped deployment.
6. What should enterprises look for when evaluating AI agent security platforms?
Look at five things: where enforcement sits and whether it blocks or only detects, where the product runs and how it is integrated, how each agent is identified and revoked, whether there is a stop mechanism and spend limits, and whether audit records reach your SIEM. Then note which features are generally available and which are in preview. The evidence that matters is a blocked call, a stopped spend, a revoked identity, and an exportable audit record.
7. Which platforms document agent identity, a kill switch, and spend limits?
IBM, Microsoft, and Prediction Guard document all three, with different scope. IBM pairs Agent Identity with a kill switch that undeploys an agent and ends its active sessions, and cost controls without a clearly established hard limit. Microsoft pairs Entra Agent ID with an organization-wide block action and Copilot Credits spending policies that are not presented as an Agent 365 feature. Blocking a Copilot Studio or Agent Builder agent affects it across Copilot and other Microsoft 365 apps, while blocking a SharePoint or Foundry agent only affects its availability in Copilot Chat; Foundry agents can be stopped separately at the Azure infrastructure level. Prediction Guard pairs a managed identity per agent with agent quarantine and enforced token and request quotas on each API key. The others do not publicly document all three.
8. Which platforms can export audit records to a SIEM?
Lakera says detection results can be exported to your SIEM, Lasso says findings can stream to one, and Prediction Guard exports audit events through an OpenTelemetry collector endpoint and log shipping. IBM says raw agent event data can be passed to existing security tooling, Microsoft documents Purview audit logs and Defender log collection, and Credo AI and Prompt Security describe audit records without publicly documenting SIEM export.
9. Where does Prediction Guard fit, and what are its limits?
Prediction Guard is a sovereign AI control plane that runs inside your boundary. It gives each agent a managed identity with an owner, limits every run to what the initiating user can reach, caps spend with per-key token and request quotas, can quarantine an agent, and exports audit events to your SIEM. Its limits are listed in the section above.