Written to be forwarded to your security team.
Regulated buyers cannot send rows to a model, and should not have to trust a vendor's word about it. Here is how the platform stays inside your control, as designed behaviour rather than policy.
No PII to LLM
Filtered at input boundary
No PCI to LLM
Blocked at source
Jailbreak detection
Template library + deny
Output scanning
Response boundary enforced
Full audit trail
App + data, every action
Air-gapped ready
Zero egress with self-hosted models
Every request is filtered before it reaches a model.
No raw input reaches a language model unchecked. The platform intercepts every prompt at the boundary, strips sensitive data, catches injection attempts, and scans what comes back, before anything is sent or returned.
PII masking
Names, emails, phone numbers, SSNs, and other personally identifiable data are detected and stripped before the prompt is formed. Nothing personal reaches the model.
Input boundary
PCI filtering
Payment card numbers, CVVs, and account details are blocked at the source. No financial identifiers leave your perimeter in any LLM request.
Input boundary
Prompt injection detection
Adversarial instructions embedded in user input or data are detected and neutralized before they can redirect or manipulate agent behavior.
Input boundary
Output scanning
LLM responses are scanned before they reach the user or downstream systems. Sensitive patterns, hallucinated PII, and policy violations are caught at the return boundary.
Output boundary
Jailbreak template library
Known jailbreak patterns and adversarial prompt templates are matched at input. Any match is denied before reaching the model — library updated continuously.
Input boundary
Global credential scanner
API keys, tokens, secrets, and credentials are detected in both input and output. Any prompt or response containing credential patterns is blocked and flagged.
Both boundaries
Deny by default
If any filter cannot confirm a request is safe, it is blocked — not passed through. The platform operates fail-closed on all prompt security controls.
Both boundaries
Platform security principles.
Schema only
The model receives table names, column names, types, and joins. Never a row, never a value. Queries run inside your perimeter against your own tables.
Filtered at the boundary
Personal data, card data, credentials, and injection attempts are caught before a request is formed, and again on the way back.
Deny by default
If a filter cannot confirm a request is safe, it is dropped and logged. Entitlement that cannot be confirmed is refused rather than guessed.
Runs where you run
Shared for a pilot, dedicated for production, or air-gapped with self-hosted models and no egress at all. Use a hosted frontier model and only the calls to that provider leave, under your policy.
Everything is recorded
Allows, denies, approvals, and handoffs land in one audit trail, each with a reason a reviewer can read.
Access scoped to the asker
Role-based access applies to every dataset and dashboard. Two people can ask the same question and correctly get different, entitled answers.
What happens to your data.
The commitments a procurement or legal review looks for — stated plainly, before you have to ask.
Data retention on downgrade
Your data is never deleted on a plan change. A grace period applies before any capability is reduced, and what we build together is yours to keep.
Snapshot-at-start
Entitlements are captured when an execution begins, so policy changes mid-run never produce surprising or unsafe behavior.
Last-known-good
If a premium check is briefly unavailable, the platform falls back to the last confirmed state for a bounded window rather than failing open.
Tenant-keyed isolation
On shared infrastructure, work and any caching are keyed to the organization and business unit — no cross-tenant leakage.
Posture & roadmap.
Responsible disclosure
Found a security issue? We want to hear from you. Email us and we'll acknowledge quickly and work with you on a fix. Please give us a reasonable window before any public disclosure.
security@nectorq.ai