Skip to main content
Get StartedOverviewPricingProofsIntegrations
The Trust Layer

Proof, not promises. Verifiable security for every LLM call.

Shield doesn't ask you to trust it. Every redaction, every tokenization, every blocked injection attempt is logged in a cryptographically verifiable hash chain, auditable, attributable, and regulator-ready from day one.

See it on your workflow
How Shield works
Security Teardown

Real threats. Real-time response.

Three common AI security scenarios and exactly how Shield handles each one, at the proxy layer, before the model sees anything.

Secret Detection
“You paste an AWS key into Claude Code”

Shield detects the AWS access key pattern before the request leaves your machine. The key is replaced with a stable placeholder, the model never sees it. Claude Code operates normally, receiving the tokenized prompt. Responses are rehydrated transparently.

PII Redaction
“A data engineer's query contains patient names”

PII patterns (names, DOBs, MRNs) are caught at the proxy layer. Patient identifiers are swapped for deterministic placeholders. The LLM processes de-identified data. Results are rehydrated with original values before the engineer sees them.

Anti-Tampering
“Prompt injection attempts to bypass filters”

Canary fingerprints embedded in every request detect tampering. If an attacker tries to strip or modify tokenized placeholders, the canary breaks. Shield blocks the request, logs the attempt, and alerts your security team via the audit sidecar.

Tamper-Evident Audit

Every redaction is cryptographically linked to its predecessor.

Shield writes every tokenization event to an append-only JSONL log. Each record includes the hash of the previous record. Modify or delete any entry and the chain breaks, immediately and verifiably.

Append-Only

Records can only be added, never modified or deleted. The log is an immutable sequence of events from Shield's first boot.

Hash-Chained

Each record carries the SHA-256 hash of its predecessor. Breaking the chain requires breaking SHA-256, computationally infeasible.

Regulator-Ready

Export the JSONL log for auditors. Every redaction is timestamped, attributable, and cryptographically linked. Satisfies PCI-DSS, SOC 2, and HIPAA evidence requirements.

Local Sidecar

The audit sidecar runs on your infrastructure alongside Shield. No audit data ever leaves your network. You control retention, rotation, and access.

🔬

See it live on your workflow

We'll deploy Shield in front of your AI tools and walk through real scenarios, paste an API key, inject PII, attempt a prompt injection, and watch Shield detect, tokenize, and log every event in real time. No slides. No promises. Just the proxy running on your machine.

Don't take our word for it. Verify it.

Shield ships with a verifiable hash chain from its first boot. Every claim on this page is backed by cryptographic evidence your auditors can validate independently. Let us show you on your own infrastructure.

Fixed-price engagement · Source code at handoff · No subscription

FAQ

Proofs & Verification Questions

How do I know the hash chain is real and not fabricated?

Every Shield audit event includes a SHA-256 hash that chains to the previous event. The genesis hash is published in the deployment architecture letter you receive at handoff. Anyone with the genesis hash can verify the entire chain independently, no access to Shield required, no trust in Purfect Labs needed. This is the same cryptographic model that powers certificate transparency and blockchain integrity.

How often are detection patterns updated?

Foundation and Compliance tiers receive quarterly pattern updates. Custom tier receives monthly updates plus emergency patches for newly discovered attack vectors. All updates ship as signed pattern packs. You control when they deploy.

What happens if Shield goes down? Does it block all LLM traffic?

Shield supports a configurable fail-mode. In 'closed' mode (default for Compliance tier), traffic is blocked if Shield is unreachable. This is the secure posture for regulated environments. In 'open' mode, traffic bypasses Shield with a logged bypass event. You choose the mode per deployment.

Is the comparison data on this page independently verifiable?

Every benchmark, detection rate, and latency figure on this page is sourced from the BENCHMARKING-PLAN.md and DETECTION-EFFICACY-PLAN.md in our public repository. We publish the exact methodology, test corpus, and measurement environment. You can reproduce every number on your own hardware.