Frequently asked questions
Blind Insight is an encryption-in-use platform: it runs queries, aggregates, machine learning, and LLM inference directly on encrypted data, using client-side keys the platform itself never holds. The answers below are the ones evaluators, security teams, and counsel actually ask — with the numbers.
The platform
What Blind Insight is, whether the math holds, and where it runs.
What is Blind Insight?
Blind Insight is an encryption-in-use platform: it runs queries, aggregates, machine learning, and LLM inference directly on encrypted data, using client-side keys the platform itself never holds. Records are encrypted by the Blind Proxy before they leave your environment; the server stores ciphertext and keyed indexes and cannot decrypt them.
Does machine learning on encrypted data actually work?
Yes. On the training algorithms Blind Insight supports — Naive Bayes, decision trees, logistic regression — the F1 delta against the same model trained on plaintext is 0.0, measured on 600,000 records. Training runs on encrypted aggregate counts, so the model learns the same decision boundaries; nothing is approximated.
Is Blind Insight hosted or self-hosted?
Both. Hosted: we run the stack and hold no encryption keys. Enterprise: we deliver a Docker container and you run everything — your data never leaves your environment. The same architecture runs identically across every plan.
Compared to FHE, TEEs, and encrypted databases
The questions evaluators ask after they've tried something else first.
What is the fastest encrypted analytics platform?
Blind Insight returns median query results in 0.12–0.9 seconds on 500,000–1,000,000 encrypted records (1–4 filters: equality, range, count, average), and trains a Naive Bayes model on 1 million encrypted records in about 50 seconds. Fully homomorphic encryption performs the equivalent training workload in 4–24 hours.
Can you train machine learning models with homomorphic encryption?
Yes, for limited models and data volumes — FHE supports the computation in principle, but at production scale it takes hours, not seconds. Duality Technologies is under active DARPA funding specifically to make FHE-based ML training faster. Blind Insight trains the same class of model on 1 million records in about 50 seconds with zero accuracy loss.
MongoDB Queryable Encryption vs. Blind Insight — what's the difference?
MongoDB QE supports equality and range queries in general availability; prefix, suffix, and substring matching are public preview, not recommended for production; full-text and fuzzy search are unsupported, and it's locked to MongoDB. Blind Insight supports equality, range, aggregates, fuzzy matching, and exact-phrase search across any structured data source.
Are TEEs (secure enclaves) safe after TEE.fail?
TEE.fail (October 2025) showed a sub-$1,000 DDR5 memory-bus interposer can extract attestation keys from Intel TDX and AMD SEV-SNP — on fully patched chips. TEEs also require specific CPU hardware, and Intel SGX is being discontinued. Blind Insight is pure software with no hardware trust dependency.
What is a privacy enhancing technology (PET)?
PETs are techniques — fully homomorphic encryption, federated learning, differential privacy, TEEs, multi-party computation, and encryption-in-use — that let organizations use sensitive data without exposing it. Blind Insight's PETs guide covers definitions of each.
Does encryption-in-use support aggregate queries like count and average?
On Blind Insight, yes — directly on encrypted fields. MongoDB QE's aggregation pipeline only groups on unencrypted fields. CipherStash's documented operators cover ordering (min/max) but no documented sum, average, or count over encrypted values as of this writing.
Is free-text or phrase search on encrypted data real-time?
Not yet for phrase search — it runs at batch and analytical speed: roughly ten seconds warm on a 10,000-record field in our own measurements, longer cold. Equality, range, aggregate, and fuzzy queries return in under a second at 500,000–1,000,000 records.
AI agents and LLMs
What an agent can reach, and why a compromised one isn't a breach.
Can an AI agent query encrypted data without ever seeing the plaintext?
Yes. Blind Insight's orchestrator sits between the agent and your encrypted store: the agent calls four tools — list schemas, describe schema, run an aggregate query, suggest an ML approach — and none of them returns a record. The agent's keyring holds only the keys its grant allows for that session — no decryption key unless a field was explicitly granted — and every call is logged.
How do you stop an AI agent from exfiltrating customer data?
We don't rely on the agent behaving. Field-level grants are scoped to what that session was approved for and revoked the moment it ends; the agent holds no decryption key for any field it wasn't explicitly granted, and Blind(L)LM's tools return aggregates, not records — so there is no raw record to exfiltrate, and every request is logged.
Is it safe to connect GPT or Claude to sensitive company data?
The model itself never touches sensitive data — it receives aggregate results computed on ciphertext, never a raw record, and it is never used for training. Blind Insight supports OpenAI, Anthropic, and Gemini models through the same orchestrator; hosted, we hold no keys, and in enterprise deployments your data never leaves your environment.
Compliance: DORA, GDPR, HIPAA
What the platform guarantees, what configuration covers, and where counsel comes in.
What compliance standards does Blind Insight support?
DORA Article 9(2), GDPR, HIPAA, and CCPA data-in-use requirements — when configured per Blind Insight's published GDPR/HIPAA/DORA compliance playbook — using FIPS-compliant AES-GCM-256 encryption. Blind Insight holds SOC 2 Type II. The guarantee underneath is simple: the server only ever sees encrypted data and never holds a key.
Is Blind Insight DORA compliant?
DORA Article 9(2) requires financial entities to protect data at rest, in transit, and in use. Blind Insight covers data in use structurally — the server only ever sees ciphertext and never holds a key. Our published compliance playbook maps controls and evidence to DORA Articles 5–9 (ICT risk management), 10–14 (incident management), 15–17 (resilience testing), 28–30 (third-party ICT risk), and 45 (secure information sharing). It isn't legal advice; your counsel signs off.
How does Blind Insight support GDPR?
The compliance playbook maps controls to GDPR Articles 5(1)(c) data minimization, 5(1)(f) integrity and confidentiality, 25 data protection by design and default, 28 processor obligations, 32 security of processing, and 33 breach-notification threshold analysis. It also names the residual risks you still own — prompt injection, output leakage, metadata linkage, and key-lifecycle failure.
Is Blind Insight HIPAA compliant?
Yes — Blind Insight is HIPAA compliant and holds SOC 2 Type II, using FIPS-compliant AES-GCM-256. Small-cohort results are suppressed automatically (k=11). The HIPAA overlay in our compliance playbook maps access, audit, integrity, authentication, and transmission-security controls, and lists the organizational controls the platform can't do for you: workforce training, sanctions, contingency planning, and BAA governance.
What k-anonymity thresholds do you recommend?
The playbook sets minimums by dataset risk tier: k of at least 5 for low risk, 10 for moderate, 20 for high. If a release doesn't meet its tier's minimum, it isn't approved — suppress or generalize the quasi-identifiers, tighten query granularity, restrict outputs to aggregates, and re-test.
Is the compliance playbook legal advice?
No. It supports implementation and documentation — control mapping, evidence packs, a 30/60/90 rollout — and says so on the page. Final interpretation and approvals belong to qualified legal counsel.
Pricing and plans
Build starts at $99/month. There is no free tier.
What counts as a record?
One indexed row inside an encrypted schema. If you're storing 50,000 patient visits across one schema, that's 50,000 records — regardless of how many fields each row has.
What happens if I hit a plan limit?
Writes throttle, reads keep working, and the dashboard surfaces an upgrade prompt. Nothing drops, nothing corrupts. You have time to decide.
Do you offer a free tier?
No. Build is the floor at $99/month. If you're working on something worth evaluating, contact us and we'll talk about the right way to try it on your own schema.
How does Enterprise pricing work?
Annual contract, anchored to API transaction volume and infrastructure footprint, with the SLA tier layered on. Contact us to scope yours.
Can I change plans anytime?
Yes. Upgrade takes effect immediately; downgrade takes effect at the next billing cycle. Your encrypted data stays where it is through the change.
Still have a question?
Book a demo and ask us directly, or try it on your own schema — nothing to deploy.