Sentinel · how it works

Proprietary framework

SENTINEL

Securing AI agents, from access to consequence.

Every permission looked fine on its own

An agent doesn't stay inside the job it was built for. Each new tool connected looked like one small, reasonable yes. By the time the combined effect is visible, it's already acted. Not recommended, acted.

An agent, not a questionnaire

What it reads, not what people remember

The usual way

Asks teams to list their agents and what each can do.

The Sentinel agent

Reads the agent platforms and permission sets themselves, including the agents nobody registered.

Step one

What it connects to, read-only

Agent platformsTool serversIdentity and permissionsCloud accountsAction logs
Sentinel agent

Read access only. Nothing is changed in your systems.

Step two

How it investigates

connectagent platforms and tool servers: read-only
discover14 agents found. 9 are registered. 5 are in no inventory.
maprefund-assistant reads CRM, payments and customer email. It can issue refunds.
testoutside input + private data + unapproved action: path open
verify71% of refund actions can be reconstructed from the logs
proposeapproval above a set value; longer log retention. Waiting for the owner.

Every line is something a person can check afterwards.

Step three

What it produces

AgentConsequenceTraceability
refund-assistantHighLow
cv-screenerHighMedium
it-helpdesk-botMediumHigh
meeting-notesLowHigh

An agent inventory and action-risk register.

How it is held to account

Rules, not vibes

01

Two scores, kept apart

What an agent can cause, and whether that can be reconstructed afterwards, are scored separately.

02

The three-ingredient test

Outside input, private data and an action nobody approves. All three together is an open path.

03

Near misses become proposals

What was learned becomes a control change for a person to approve.

The agent is held to the same questions it asks of everyone else.

Configured to your regulator

The same agent. Your market.

SM&CR accountability and PRA expectations decide who must own each agent and what must be reconstructable.
FINMA's governance expectations and the Federal Act on Data Protection decide the evidence and retention rules.
The EU AI Act's logging and human-oversight duties, with DORA's resilience duties, decide the evidence rules.

Choose a market. Other markets are added as packs.

How an engagement runs

From first conversation to a working agent

1

Scope

Which platforms, which business units, who owns each agent.

2

Connect

Read-only connections, reviewed by your security team first.

3

First register

The agent inventory and the action-risk register.

4

Keep watching

New agents and changed permissions are flagged as they appear.

A typical shape. The sequence and timings are set with you.

One platform

Sentinel does not work alone

KeelSets the policy and the gates for every use case.
SentinelQuestions what each agent can see, change and cause.
SoundingsSees what is actually running across the organisation.
BearingProves it to the supervisor, obligation by obligation.
The human layerHALO, Squall and Ensign measure what no system log can.

See the whole system on the Solutions page.

That's Sentinel

Want to know what your AI agents can actually do?

A demonstration of the method is one click away. The deployed Sentinel is built for your environment and configured to your regulator.

Discuss a deployment