PRODUCT / TRAINING DATA AUDIT

Find poisoned samples before you train.

250 documents are enough to backdoor an LLM. 32 examples are enough to blind-spot a guardrail. We scan your dataset before training and show you exactly which samples to remove.

// Project-based pricing. No subscription. CPU-only. EU sovereign.

The backdoor is in the data, not the model.

When a model is poisoned, the backdoor is baked into the weights. Runtime firewalls cannot see it — they detect symptoms, not the cause. To stop a training-time backdoor, you have to find it BEFORE training starts.

250

documents are enough to backdoor an LLM of any size. This is a fixed count, not a percentage — invisible in a dataset of 260 billion tokens.

ANTHROPIC / UK AISI / TURING (OCT 2025)
32

poisoned examples are enough to blind-spot a Constitutional Classifier. It retains 99.9% accuracy in standard tests.

CONSTITUTIONAL CLASSIFIER POISONING (APR 2026)
3-5%

poisoned preference data is enough to introduce a backdoor. The effect is log-linear — more poison, stronger backdoor.

POISONBENCH (ICML 2026)

The EU AI Act (Article 15) requires high-risk AI systems to be resilient against data poisoning. Fine-tuning without an audit is a compliance risk.

What we scan for.

Nine detectors run against your dataset before training. Each one targets a different class of poisoning.

DETECTOR 01

RARE N-GRAMS

Compute n-gram frequency. Flag rare n-grams with high label correlation — potential triggers.

STATISTICAL CORRELATION
DETECTOR 02

NEAR-DUPLICATES

MinHash deduplication. Flag clusters of near-identical samples — deduplication poisoning.

MINHASH LSH ALIGNMENT
DETECTOR 03

EMBEDDING CLUSTERS

DBSCAN on sentence embeddings. Flag small, isolated, high-purity clusters — poison clusters.

DBSCAN TOPOLOGY
DETECTOR 04

PERPLEXITY ANOMALIES

Per-sample perplexity via a small base model. Flag samples that spike perplexity — DoS payloads.

CROSS-ENTROPY THRESHOLD
DETECTOR 05

HIDDEN UNICODE

Detect zero-width characters, Tag sequences, BOM, soft hyphens, direction overrides.

CODEPOINT BYTE PARSER
DETECTOR 06

ENCODED PAYLOADS

Decode Base64, hex, ROT13, Caesar, Morse. Flag samples where decoded content contains instructions.

20+ MULTI-CIPHER DECODERS
DETECTOR 07

LABEL FLIPS

Detect harmful content labeled as safe. Works for SFT and preference format (chosen/rejected).

ALIGNMENT INVERSION
DETECTOR 08

MUTATED TRIGGERS

Levenshtein clustering on rare n-grams. Flag almost-backdoor variants — attacks that bypass keyword filters.

FUZZY EDIT DISTANCE
DETECTOR 09

JSON + PDF REPORTS

Per-sample risk score, aggregate risk (0-100), trigger candidates, cluster analysis, recommendations.

STAKEHOLDER COMPLIANCE PDF

How it works.

Upload your dataset. We scan it. You get a report with per-sample risk and remediation guidance.

01

UPLOAD

Send your dataset via API, or point us to an S3/HuggingFace location. Supports SFT format and preference format (prompt/chosen/rejected).

FORMAT: JSONL / CSV / PARQUET
02

SCAN

Nine detectors run in parallel. Async jobs for datasets up to 50k samples. On-prem option available for sensitive data.

ASYNC CPU WORKER QUEUE
03

REPORT

JSON report (machine-readable) and PDF report (for stakeholders). Per-sample risk score, trigger candidates, cluster analysis, recommendations.

PER-SAMPLE RISK + PDF ARTIFACT
04

REMEDIATE

Remove flagged samples, re-scan, confirm clean. We document exactly which samples to remove and why.

ZERO-POISON CERTIFICATION
API WORKFLOW // ASYNCHRONOUS AUDIT SUBMISSION (cURL EXAMPLE)
# Example Dataset Audit Job Submission (cURL)
curl -X POST https://synaptosec.com/v1/training-audit/scan-async \
 -H "Authorization: Bearer <YOUR_API_KEY>" \
 -H "Content-Type: application/json" \
 -d '{
 "texts": ["sample document 1", "sample document 2"],
 "labels": ["train", "train"],
 "firewall_enabled": true
 }'

// CLI toolkit & CI/CD runner: coming soon. Validation results: coming soon.

COMMERCIAL MODEL // PROJECT-BASED

Project-based pricing. Zero subscription.

Designed for ML Engineers, Heads of AI, and Enterprise Risk Teams. Pay per scan or per fine-tuning run without ongoing seat licenses.

PROJECT TIER 01

Single Dataset Audit

Up to 10,000 samples. Fast validation before production fine-tuning.

  • All 9 parallel detectors
  • Machine-readable JSON export
  • Executive PDF audit certificate
  • 24-hour turnaround
MOST COMMON
PROJECT TIER 02

Enterprise Batch Scan

Up to 50,000 samples. Full SFT or preference corpus with remediation re-scan.

  • Includes 1 complimentary remediation re-scan
  • DBSCAN embedding poison cluster graph
  • EU AI Act Article 15 compliance pack
  • Dedicated ML security engineer support
PROJECT TIER 03

Air-Gapped / On-Prem

Run locally inside your private VPC or sovereign data center. Zero egress.

  • Docker container / Kubernetes Helm chart
  • Zero outbound network telemetry
  • CPU-only inference (no GPU required)
  • Classified & regulated banking support

Secure your training pipeline.

Send us a dataset sample or discuss air-gapped on-prem deployment with our security engineers.

// Built for ML Engineers & AI Security Leads. Designed to support GDPR and EU AI Act Article 15 compliance. Validation results: coming soon.