Inference guardrails

Pipelock vs Lakera Guard

Inspection of routed network traffic next to a classifier API at the model boundary, now sold by Check Point as AI Guardrails.

At a glance

Pipelock source Lakera Guard
Job Agent firewall. Mediates HTTP, WebSocket, and MCP traffic routed through it, scans it for secret leaks, prompt injection, SSRF, and tool poisoning, and can emit signed action receipts for mediated decisions when a signing key is configured. Classifier API at the model boundary. Screens prompts and completions for injection, jailbreaks, data leakage, unsafe tool use, and harmful content.
Enforcement point Network path, outside the agent process Between the application and the model; the application calls the guard API per interaction
Source Open source, Apache-2.0 core; Enterprise under ELv2 Closed source (Check Point)
Pricing shape Free core; paid Pro and Enterprise tiers Free tier and demo on the site; self-hosting needs an Enterprise license
Runs as Single Go binary, self-hosted; container and Helm SaaS by default; self-hosted via Helm, Docker, or air-gapped containers with an Enterprise license
Pick Lakera Guard

Your application is a chatbot, RAG pipeline, or copilot and the risk you care about is what goes into and comes out of the model.

Pick Pipelock

Your agent takes actions after the model answers, and you need traffic routed through Pipelock inspected on the network path and recorded in a form an auditor can verify.

Run both

Lakera screens submitted model interactions. Pipelock inspects the agent traffic routed through it after the model response.

Want the runtime boundary, not just another checklist?

The short version

Pipelock is an open-source agent firewall. It scans HTTP, MCP, and WebSocket traffic routed through it for credential leaks, prompt injection, SSRF, and tool poisoning. With a signing key configured, it can emit signed action receipts for mediated decisions. Single Go binary, self-hosted.

Lakera Guard is a runtime classifier for LLM applications. Check Point acquired Lakera, and the product page now carries the AI Agent Security and AI Guardrails names. The integration shape is a call to the guard API on each model interaction: the application sends the input, optionally the model’s response, and gets back verdicts for prompt injection, jailbreaks, data leakage, unsafe tool use, harmful content, and multilingual or multimodal attacks. SaaS is the default; the docs describe self-hosting via Helm, Docker, or air-gapped containers under an Enterprise license.

Pipelock inspects traffic the agent routes through it. Lakera classifies content the application sends to its API.

Feature comparison

FeaturePipelockLakera Guard
ArchitectureNetwork proxy, single self-hosted binaryClassifier API; SaaS with an enterprise self-hosted option
Enforcement pointMediated agent egress: HTTP, HTTPS CONNECT (payloads only with TLS interception on), WebSocket, MCPSubmitted LLM interactions and tool calls
HTTP egress scanningMediated HTTPNot a proxy; screens content the application submits
WebSocket frame scanningMediated WebSocket framesNot a proxy; screens content the application submits
MCP awarenessMediated tool descriptions, arguments, and responsesScreens submitted tool calls and separately submitted tool definitions; drift detection not documented
Credential scanning (DLP)65 built-in patterns with encoding normalization on routed trafficData-leakage classification on submitted content
Prompt injection detectionDeterministic patterns with multi-pass normalizationModel-based classifier
Tool poisoning and rug-pull detectionMediated tool-description scanning and drift detectionTool-definition screening is documented; drift detection is not documented
SSRF and private-IP blockingMediated destination checks before DNS resolutionNot documented
Latency claimScan latency around a millisecond on the mediated path“Delivering sub-50 ms runtime latency”, per the product page
Integration callNone; the agent’s proxy setting or host containment routes traffic through itOne POST to /v2/guard per interaction, returning a flagged boolean and an optional per-guardrail breakdown
Observe-only modeAudit mode and per-scanner warn actionsDetect mode per project, which forces flagged to false while logging
Per-decision evidenceSigned receipts when configured, plus a hash-chained flight recorder; verifiable offlineVerdict in the API response; the application logs it; a results endpoint for historical analysis
Compliance mappingsOWASP MCP Top 10, OWASP Agentic Top 15, OWASP LLM Top 10, NIST 800-53, EU AI ActFramework coverage page in the docs
Process sandboxLandlock, seccomp, and network namespaces on LinuxNo
Source availabilityApache-2.0 core; Enterprise under ELv2Closed source
Self-hostedYes, standard deploymentEnterprise license required

When to pick Pipelock

Your threat model extends past the model call. Agents that fetch pages, open WebSockets, or POST to APIs take actions that a classifier API sees only if the application submits those contents. A leaked cloud key in a request body does not automatically traverse that API. Pipelock scans that traffic when the agent routes it through Pipelock.

You need MCP transport mediation. Pipelock scans and fingerprints tool descriptions on its MCP path, then inspects arguments and responses. Lakera can screen submitted tool calls and tool definitions, but its docs do not describe MCP transport mediation or drift detection.

You want deterministic, reproducible enforcement and offline-verifiable evidence. Pipelock’s patterns are open source, so a match is reproducible from the same input. With a signing key and allow-path receipt enforcement configured, it blocks an allowed mediated request rather than forward one whose signed receipt cannot be emitted. The shipped verifier checks signed receipts offline.

You want to self-host the Apache-2.0 core. Pipelock runs as a self-hosted Go binary.

When to pick Lakera Guard

Your application is a chatbot, RAG pipeline, or copilot. The control point that matters is the model boundary, and one API call per interaction is a clean integration.

You want a classifier API at the model boundary. The Guard API supports Detect and Enforce project modes, so the application can monitor its policy before choosing a blocking response.

You already run Check Point. Vendor consolidation is a real consideration for a security team on that platform.

You need sub-50 ms classification. That’s the figure on the product page for Lakera Guard.

Architecture differences

Lakera Guard asks: does the submitted LLM interaction violate policy? The application can POST an interaction to /v2/guard before or after a model call, then choose how to handle its flagged verdict. In Detect mode the verdict is always false and the result is logged. The docs also tell integrators to pass system instructions as separate system-role messages so they are not flagged as prompt attacks. Content reaches the guard only when the application submits it.

Pipelock asks: what is moving across its mediated agent boundary, and is any of it unsafe? Routed bodies are checked for leaked credentials, mediated tool descriptions for poisoning and drift, scanned responses for injection, and SSRF is checked before DNS resolution. With a signing key configured, it can emit signed action receipts for those mediated decisions.

Lakera screens submitted model interactions. Pipelock inspects traffic routed through it after the model response. Run both when both layers matter.

A note on evidence

Lakera returns a verdict in the API response. The application decides whether to log it and holds the record.

With a signing key configured, Pipelock can emit signed receipts for mediated decisions. Each carries the action, policy hash, decision, and a hash link to the previous receipt. With require_receipts enabled, a failed allow-path receipt emission blocks forwarding. The shipped verifier checks a chain offline against the deployment’s public key without contacting a Pipelock service, the agent, or the application. The action receipt format is public. It’s the operator’s key, so it proves what the mediator decided; it doesn’t make the operator independent.

Further reading

Sources checked

Third-party descriptions on this page come from the public materials below, read on the dates shown. Features and pricing change; check the current documentation before you decide.

  • Lakera Guard product page checked 2026-09-01 · now branded as Check Point AI Agent Security; states sub-50 ms runtime latency
  • Lakera docs: self-hosting AI Guardrails checked 2026-09-01 · Helm, Docker, and air-gapped options; requires an AI Guardrails Enterprise license
  • Lakera docs: guard API endpoint checked 2026-09-01 · POST /v2/guard on api.lakera.ai; flagged boolean; Enforce and Detect project modes; optional breakdown and payload fields; guard/results endpoint for historical analysis; page is branded Check Point AI Guardrails

Third-party product names and marks belong to their owners. PipeLab is not affiliated with, sponsored by, or endorsed by the makers of any product compared on this page. Descriptions of other products come from their own public materials on the dates listed above and reflect PipeLab's reading of them. If something here is wrong or out of date, tell us and it will be corrected.

Frequently asked questions

What's the difference between Pipelock and Lakera Guard?
Pipelock is an open-source agent firewall that scans HTTP, WebSocket, and MCP traffic routed through it for credential leaks, injection, SSRF, and tool poisoning. With a signing key configured, it can emit signed action receipts for mediated decisions. Lakera Guard is a classifier API, now part of Check Point, that screens submitted model interactions. Different enforcement points, different evidence.
Does Lakera Guard scan MCP or HTTP egress traffic?
Its Guard API screens LLM interactions that the application submits. Its documentation recommends calling it on every interaction and tool call, and says tool definitions are screened separately. It is not an HTTP egress proxy, so a request, WebSocket, or MCP exchange reaches Lakera only when the application sends that content to its API. Pipelock can inspect those transports when they are routed through Pipelock.
Can I use Pipelock and Lakera Guard together?
Yes. Lakera protects the model conversation. Pipelock protects the agent’s actions downstream of the model. A deployment that cares about both layers runs both.

Want the runtime boundary, not just another checklist?

See all comparisons →