Security and audit

Every message checked. Every step on record.

Guardrails run on your server before and after the model. Personal data is filtered, and an audit log your auditor can verify records every step.

Agent StudioGuardrailsevery message
Inbound
✓
Input lengthbefore the model
✓
Prompt injectionbefore the model
✓
Deny listbefore the model
✓
Personal databefore the model
Outbound
✓
Prompt protectionbefore the customer
✓
Personal-data scanbefore the customer
✓
Safety deny listbefore the customer
✓
Reply lengthbefore the customer
Protection

Four layers, one click apart.

Sixteen filters before the model sees a word.

Contact, identity, financial and health identifiers become placeholders before the model, the history or the logs.

Agent StudioPersonal-data filters16 built in
Email[EMAIL]
US phone[PHONE]
US SSN[SSN]
Pakistani CNIC[CNIC]
UK NI number[NINO]
IBAN[IBAN]
Payment card[CARD]
Medicare ID[MBI]
Date of birth[DOB]
Medical record no.[MRN]
Claim number[CLAIM]
Provider NPI[NPI]
+ up to 25 custom filters per agent

Card numbers removed, typed or spoken.

Caught even when read out as "four one double one", checked with the Luhn formula, and replaced before storage. Payments go through a payment link.

  • Designed to keep the agent out of PCI DSS scope. Your assessor confirms.
Agent StudioGuardrailswhat the model receives16 filters on
Hi, my SSN is 512-48-3907[SSN]. Was claim CL-88213[CLAIM] paid?
My card is 4111 1111 1111 1111[card number removed]
Refund to GB82 WEST 1234 5698 7654 32[IBAN]

The log proves what happened without storing what was said.

Append-only, in its own database, chained with SHA-256. Edits and deletes are refused even for the owner.

  • One-click verify pinpoints a missing or changed entry.
  • CSV export, protected against formula injection.
Agent StudioAudit loglending✓ chain verified
✓call.received a91f…3c
✓pii.redacted SSN, DOB7be0…a1
✓tool.validated 23 rulesc40d…9e
✓customer.confirmed £148.2018e2…f7
✓workflow.completed e7a9…02

Tools that cannot go rogue.

Deny-by-default tool lists, server-side validation, customer confirmation and a call budget, then a network guard on every outbound call.

Agent StudioNetwork guardevery API tool
https://api.corebank.exampleAllowed
http://169.254.169.254/Refused · metadata
http://10.0.4.12/adminRefused · internal
http://0x7f000001/Refused · disguised
302 → elsewhere.exampleRefused · redirect
Isolation

Three stores, kept apart.

Every team's records are isolated. Ask for another team's record and the answer is "not found".

Platform database

Users, teams, roles and workflows. Connector credentials encrypted with AES-256-GCM.

AI service database

Agents, masked conversations and knowledge. Tool evidence encrypted with AES-256-GCM.

Audit database

Append-only and hash-chained. Edits and deletes are refused.

Compliance

Controls designed to support HIPAA, SOC 2 and PCI DSS programmes.

Compliance belongs to your organisation. These are the controls Alphanetix gives your programme to build on.

AlphanetixControl mapping
Control areaWhat Alphanetix providesSupports
Access controlTeam-scoped requests; masked history revealed only with permission.HIPAA · SOC 2
Audit controlsAppend-only log of tool calls, reveals, exports, deletions and retention changes.HIPAA · SOC 2
IntegritySHA-256 hash chain, verified in one click.HIPAA · SOC 2
AuthenticationStudio sign-in; revocable access keys bound to each conversation.HIPAA · SOC 2
Transmission securityEncrypted voice transport; unencrypted connections refused in production.HIPAA
Cardholder dataCard numbers removed before storage, the model and the logs.PCI DSS
Talk to our team

Bring your security questionnaire.

We will walk your security team through each control, on your own infrastructure.