The rows worth reading.
Lakera is a security tool: its core product, Lakera Guard, detects prompt injection, jailbreak attempts, and PII leakage — known categories of adversarial or unsafe content. It has no concept of your business policy, so it has nothing to say when an agent confidently, safely, and incorrectly approves a refund outside your return window or quotes the wrong price. Zespan covers that gap: guardrails written against your actual rules, backtested against real traffic, enforced in-process, with approval holds and an audit trail for the decisions a security scanner was never built to catch.
| Zespan | Lakera | |
|---|---|---|
| Prompt injection / jailbreak detection | Via custom guardrail rules | Core feature |
| PII and secret leakage detection | Built-in guardrail type | Core feature |
| Catches a well-formed but policy-violating decision | Rules written against your policy | Not available |
| Catches a wrong tool-call argument | Not available | |
| Backtests a rule on your own past traffic first | Catch rate + false-positive rate | Not available |
| In-process enforcement, no proxy hop | SDK evaluates locally | Proxy/API-based |
| Approval holds for risky actions | Named reviewer, recorded decision | Not available |
| Full agent tracing (spans, cost, delegation) | Not available | |
| Audit trail mapped to compliance controls | Actor-attributed, SOC 2-mapped | Not available |
| Starting price | $29/month | Contact for pricing |
Which one fits
Pick Zespan when
Your agent takes real actions — refunds, approvals, quotes — where a wrong decision costs money even though it's perfectly safe content.
You need a rule backtested against your own traffic before it can block or approve anything.
You want tracing, cost attribution, and evaluations in the same platform as your guardrails, not a separate tool.
You need approval holds and audit evidence for actions above a risk threshold.
You want guardrails enforced in-process, without routing every request through an external proxy.
Pick Lakera when
Your primary exposure is adversarial input — public-facing chat exposed to untrusted users — and prompt injection is your top concern.
You want a specialized, deep security-scanning product and are handling business-logic correctness separately.
You need Lakera's specific threat-intelligence dataset and detection models for adversarial content.
Common questions
Can I use Zespan and Lakera together?
Yes, and for agents exposed to untrusted or adversarial input, many teams do: Lakera for adversarial content detection, Zespan for the business-logic correctness and enforcement layer Lakera doesn't cover, on the same production agent.
Does Zespan catch prompt injection like Lakera does?
Zespan's guardrail system supports rules for known-bad content categories including PII and secret leakage as a built-in type, and prompt-injection-style rules can be authored as custom guardrails. Lakera's core product is purpose-built and specialized specifically for adversarial content detection at a depth Zespan doesn't claim to match.
Why doesn't a security guardrail catch a wrong refund approval?
Because nothing about a wrong refund approval is adversarial or unsafe. The response is well-formatted, on-topic, and safe. It's just incorrect against your specific refund policy — a business-logic violation, not a security one, and a different detection problem entirely from prompt injection or PII leakage.