Sectum AI vs Lakera

TL;DR. Lakera is an AI-native security platform (Guard for runtime protection, Red for adversarial testing, PII Detection for sensitive-data filtering) that sits in the request path and blocks or scores AI traffic. Sectum AI sits across the multi-tenant boundary and verifies and attests — it provisions synthetic tenants, plants cryptographic markers, runs cross-tenant probes across 13 surfaces, and produces a tamper-evident evidence pack. The two products solve different problems for the same buyer and complement each other; neither replaces the other.

The two products

Lakera (lakera.ai)

Category: AI-native security platform. Three products:

License: commercial. No open-source core.

Pricing (Lakera platform pricing):

Public customer: Dropbox uses Lakera Guard as an in-VPC Docker microservice for AI-powered features (smart search, document summarization).

Funding: Series A; specifics for 2026 not publicly surfaced.

Sectum AI (sectum.ai)

Category: multi-tenant AI verification. Not a runtime guardrail. Not a firewall. The deliverable is auditor-acceptable, tamper-evident evidence that the tenant boundary holds across the AI stack.

License: Apache 2.0 OSS core (substrate, attack catalog, adapters, evidence chain, sectum-ai verify). Sectum Cloud commercial. The evidence layer in the OSS produces the same artifacts the hosted product does — by design.

Method: synthetic-tenant marker substrate. Provisions tenants on the customer’s AI stack, seeds them with cryptographic canary markers and a hashed ground-truth manifest, runs 11 cross-tenant probe classes across 13 surfaces, produces a tamper-evident evidence pack (RFC 3161 TSA + Sigstore Rekor + in-toto envelope, control-mapped audit PDF, machine-readable evidence.json).

The categorical difference: blocking vs. verifying

The fundamental difference is what the product does to a request:

Lakera GuardSectum AI
ModeRuntime: intercept, score, and block trafficPeriodic verification: provision synthetic tenants, run probes, produce evidence
PositionIn the request pathOutside the request path
UnitPer-request decision (allow / block / score)Per-run audit pack
DetectionReal-time content filtering, prompt-injection classifierManifest-grounded layered detection (exact → semantic → calibrated judge)
OutputBlock / allow / score per requestTamper-evident audit pack (RFC 3161 + Rekor + in-toto + PDF + JSON)
ForApplication engineering and platform securityCISOs, DPOs, audit firms
DeploymentAPI or in-VPC Docker microservice (request path)CLI in customer’s environment (BYOC); only signed evidence leaves
When it firesEvery requestOn a schedule, on-demand, or at every audit cycle / Article 17 ticket
What it can attestLive blocking rate, per-policy metrics”Tenant A’s data cannot reach tenant B across these 13 surfaces under this manifest hash, signed and cryptographically timestamped”

Lakera Guard is a control — a runtime mitigation that does work on every request. Sectum AI is a verifier and attester — it doesn’t touch live traffic; it produces evidence that the boundary holds. Both are useful; they answer different questions; deploying both on the same stack is the natural pattern for a serious multi-tenant AI SaaS.

Lakera Red and Sectum AI: closer, still different

Lakera Red is the closest thing in Lakera’s portfolio to Sectum AI — both test the AI system rather than blocking traffic. But the unit of analysis differs:

If you want broad adversarial testing of an AI system, Lakera Red is a strong commercial option. If you specifically need to prove multi-tenant isolation with auditor-grade evidence — and if GDPR Article 17 erasure attestation is on your roadmap — Sectum AI is the focused tool.

Surface coverage

SurfaceLakera GuardLakera RedSectum AI
LLM endpoint (input/output filtering)✓ (runtime)✓ (test)✓ (probe surface)
PII detection in real time— (not a goal — Sectum AI tests cross-tenant flow, not PII filtering)
Cross-tenant boundary on a shared vector DBpartial (general probes)✓ (Class 2 + direct Pinecone/pgvector/Weaviate/Chroma adapters)
Semantic-cache contamination✓ (Class 4 + live Redis adapter)
KV-cache timing side channel✓ (Class 5, statistical effect-size test)
Embedding inversion across tenants✓ (Class 6)
Agent / MCP confused-deputy + token passthroughpartial (Lakera Guard for agents)partial✓ (Class 7 — the Asana-class flaw with per-finding evidence)
Persistent agent memory cross-tenant✓ (Class 8)
LoRA / fine-tune cross-tenant influence✓ (Class 9)
Multi-turn benign extraction (IKEA/Silent Leaks)partial✓ (Class 10)
RAG poisoningpartial✓ (Class 3)
GDPR Article 17 erasure verification✓ (Class 11 — the Erasure Attestation engagement)
Observability backends (Langfuse/LangSmith/Phoenix)— (Lakera has its own observability)✓ (live adapters; erasure verifies these too)

Lakera Guard is depth on the request path (blocking, filtering, real-time PII). Sectum AI is depth on the tenant boundary across 13 surfaces. The two coverages are perpendicular.

Evidence model

Lakera Guard’s output is per-request decisions and platform telemetry. Lakera Red’s output is a vulnerability report. Both are excellent for security operations; neither is shaped like an auditor attestation.

Sectum AI’s output is a different artifact:

For an auditor or DPO asking “can you prove tenant A’s data didn’t reach tenant B?” — the Lakera platform gives a runtime story; Sectum AI gives a cryptographic chain of custody.

When to use Lakera

When to use Sectum AI

Using both

The mature multi-tenant AI SaaS deploys Lakera Guard in the request path (runtime protection) and runs Sectum AI periodically (verification + auditor evidence). They serve different parts of the same security posture:

Neither blocks live traffic the way Lakera Guard does. Neither produces the cryptographic chain of custody Sectum AI does. The two compound, and using both is the cleanest pattern for an AI shop that takes both runtime protection and audit readiness seriously.

The “AI security” category, broken down

“AI security” is a label that mixes runtime guardrails, adversarial testing, supply-chain scanning, and audit evidence — all under one banner. A more precise breakdown:

A serious AI program touches several of these. Sectum AI focuses on the verification and attestation slice at the multi-tenant boundary.

Pricing

References


← All comparisons