Enterprise AI governance · Built on Aegis

Every agent action. Accounted for.

Anodyne enforces AI policy in the path of every model call, at the edge and inside your cloud, and keeps an auditable record of each one.

Request access How it works

In the path

Two gates, one policy. Enforcement happens inline, before the model.

Your rules, per workload

One policy decides what is allowed, what is recorded and how long it is kept.

On the record

Every call is recorded and queryable: which gate, which model, what it cost.

The gap

Your agents already have production access. Your governance still has a quarterly review.

Delegation

Agents act on borrowed credentials. They inherit a person's access and run unattended.

Policy

Policy lives in documents. It is absent from the execution path, where actions happen.

Audit

Audit happens after the fact. Logs explain an incident. They do not prevent one.

Why now

The actor changed.

For twenty years, enterprise security assumed the actor on the other side of a permission was a person. That assumption is ending.

Agents act

Agents now read the CRM, write to the warehouse, open pull requests and move money, with credentials delegated by people who are not watching.

Documents cannot keep pace

Governance written in documents cannot keep up with software that acts in milliseconds. The control has to live in the path of execution, decide before the action happens, and leave a record a regulator can read.

Filters alone do not hold

Model-based filters reduce risk but yield under adversarial pressure. The boundary must be identity and network; screening is a signal; the record is the backstop.

Agents are the new insiders. Govern them like it.

What it catches

The mistakes that cost money. Stopped before they happen.

Our own agents take about 3,500 actions a day, and every one is recorded, attributed and costed. These are the slips the same controls are built to catch.

A model nobody approved

Every gate carries its own allowlist, so a request for a model off the list is refused.

Why it mattersTeams cannot quietly adopt an unvetted model. Procurement and security decide which models are in use.

A spend spike

Daily spend and call ceilings cap each gate, and budget alerts go out by business domain.

Why it mattersThe AI bill stops surprising the CFO. Every dollar is attributed to a business domain.

A call that skips the gate

There is a single enforced path to every model, so there is no side door to leave open.

Why it mattersThe policy you wrote is the policy that runs. No workload can route around it.

A secret in a prompt

Secrets and personal data are redacted inline, before the prompt leaves your boundary.

Why it mattersA pasted credential or customer record does not become a breach notice.

A call nobody can account for

Each model call leaves a record you can query in SQL, with a cloud audit trail behind every private call.

Why it mattersWhen the board or a regulator asks what the AI did, the answer is a query, not a reconstruction.

A rule that quietly changed

Policy is code, versioned per gate, and both gates only run signed code.

Why it mattersChanges to the rules are reviewed and traceable, not edited in place.

Anodyne is a guard against mistakes, not a boundary against a determined insider. Every capability below is marked Live, Building or Planned.

Architecture

Policy in the path, not in a PDF.

Approved workloads reach AI providers through the edge gate. Restricted workloads never leave your cloud: they run through the private gate. Both enforce one policy, every call is recorded, and no other path to a model exists.

Agents

Copilots, assistants, automations and the tools they call

Yours

Anodyne

One policy, enforced on every call

Edge gate Live

For approved AI providers

Private gate Live

Inside your cloud, nothing leaves

Live

Aegis

Security, records and cost controls underneath, built and changed only through reviewed code

Live

Models and systems

AI providers, models hosted in your cloud, your data

Yours

Fig 1. Observability tells you what happened. Governance decides what is allowed to.

Your posture

Three questions. Eight known setups.

Every workload is placed by three decisions. Each combination maps to a defined configuration with known controls and known evidence.

Question 1

Where may it run?

OutsideApproved AI providers, via the edge gateInsideYour cloud only, via the private gate

Question 2

How closely is it watched?

GuardedScreened before reaching the modelRecordedRedacted, capped and logged

Question 3

How strict are the rules?

StandardDefault retentionRegulatedExtended retention, restricted models

2 × 2 × 2 = eight setups, each with known controls and known evidence.2 live5 planned1 not offered

Live

Edge gate, guarded

Approved AI providers through the edge gate. Every call redacted, screened, capped and recorded.

Controls

Edge gateLive
Model allowlist per gateLive
Inline redaction of secrets and personal dataLive
Content screening, monitor mode (edge)Live
Daily spend and call ceilingsLive
Per-call records, queryable in SQLLive

Planned

Edge gate, guarded, regulated

The edge configuration with extended retention and a restricted model list.

Controls

Edge gateLive
Inline redaction of secrets and personal dataLive
Content screening, monitor mode (edge)Live
Per-call records, queryable in SQLLive

Still to build

Compliance profiles per client
Profile-driven retention
Tamper-evident records

Planned

Edge gate, record only

For low-risk tooling: redacted, capped and recorded, without screening.

Controls

Edge gateLive
Inline redaction of secrets and personal dataLive
Daily spend and call ceilingsLive
Per-call records, queryable in SQLLive

Still to build

Per-workload screening control

Not offered

Regulated data does not leave your cloud unscreened. Use the private gate or add screening.

Planned

Private gate, guarded

The private gate with pre-model screening, introduced in monitor mode.

Controls

Private gateLive
Single enforced path to every modelLive
Per-call records, queryable in SQLLive
Cloud audit trail for every private callLive

Still to build

Content screening (private gate)

Planned

Private gate, guarded, regulated

Maximum assurance: data stays in your cloud, every call is screened, and records are retained to your requirements.

Controls

Private gateLive
Single enforced path to every modelLive
Per-call records, queryable in SQLLive
Cloud audit trail for every private callLive

Still to build

Content screening (private gate)
Compliance profiles per client
Tamper-evident records
Authenticated caller identity

Live

Private gate

Data never leaves your cloud. A single enforced path, signed code, redaction, ceilings and a complete record of every call.

Controls

Private gateLive
Single enforced path to every modelLive
Model allowlist per gateLive
Signed-code enforcement on both gatesLive
Daily spend and call ceilingsLive
Per-call records, queryable in SQLLive
Cloud audit trail for every private callLive

Planned

Private gate, regulated

The private gate with extended, protected retention for regulated workloads.

Controls

Private gateLive
Single enforced path to every modelLive
Per-call records, queryable in SQLLive
Cloud audit trail for every private callLive

Still to build

Compliance profiles per client
Profile-driven retention
Tamper-evident records

Capabilities

Prevent. Detect. Prove.

No single filter stops a determined attacker. Anodyne layers hard limits, screening and a record, so when one layer misses, the next one catches it.

Prevent

Controls that hold regardless of model behavior.

Single enforced path to every modelLive
Model allowlist per gateLive
Inline redaction of secrets and personal dataLive
Daily spend and call ceilingsLive
Signed-code enforcement on both gatesLive
Policy as code, versioned per gateLive
Compliance profiles per clientPlanned

Detect

Screening calibrated in monitor mode before it blocks.

Content screening, monitor mode (edge)Live
Budget alerts by business domainLive
Content screening (private gate)Planned
Category blocking with an explicit fail modePlanned

Prove

A complete record of every call, held in the gate’s own cloud account.

Per-call records, queryable in SQLLive
Cloud audit trail for every private callLive
Cost attribution by business domainLive
NIST 800-53 control mapping, verified on every changeLive
Tamper-evident recordsPlanned
Authenticated caller identityPlanned

Every control is labelled by what it does today. Recorded is never presented as blocked.

By role

One record. Three readers.

CISO

Hold the boundary, not the roadmap.

Every path to a model runs through a gate, enforced in cloud permissions. Teams ship agents; security holds the boundary and an audit record the board can rely on.

Example gate policy

gate     private
models   [approved list]
mask     secrets, personal data
cap      [n] calls a day
no record DENY

CIO

AI spend, attributed and capped.

Spend by model, gate and business domain. Daily ceilings stop runaway agents; budgets alert before month end.

Example policy

domain   [business domain]
budget   $[n] a month
cap      $[n] a day, edge gate
alert    at [n]% of budget

CDAO

Every model call has a record. Now you can query it.

Gate, model, tokens, cost and redactions for every call, queryable in SQL and retained to your requirements.

Example record

gate     private
model    [approved model]
tokens   in [n] · out [n]
masked   [n] fields
kept     [your period]

08 · Request access

Start with one workload.

Tell us where your agents run today. We will place one workload, put a gate in front of it and deliver the record.

We use what you send to respond, and keep it 12 months. We will never use cookies to track you. The form uses Cloudflare Turnstile to keep out bots. Privacy notice

connect@anodyne.app