Security for the AI agents you run

Understand what your AI agents can reach and put controls around
that access. An agent connected to mail, documents or an API
introduces permissions and failure paths that need review. We
inventory the agreed systems, assess the risks, implement approved
controls and test the restricted actions. You receive the access
map, findings and verification record.

Mapped: every agent and what it can reach

We inventory agents within the agreed scope, including approved
integrations and other connections identified with your staff. For
each, the map records credentials, tools and reachable data. You
receive the map with the findings and responsibility for keeping it
updated is agreed as part of handover. It is a record of the
reviewed environment, with its scope and gaps made explicit.

Locked: each agent on least privilege

An agent gets its own identity and the smallest privilege its job
needs: roles rather than shared accounts (RBAC), read-only by
default, writes only through narrow tools with limits on how often
and how much. Secrets are split per agent and per host, never pasted
into a prompt, and rotated without a code change. External text an
agent reads — an email, a web page or a document — is treated as
untrusted data. Tool permissions and validation checks enforce the
agreed boundaries; a message saying "ignore your rules" is not
authorization for an action. Controls fail closed: when a check
cannot run, the action is refused.

Recorded: what it did, and for whom

Every tool call is logged with the person it acted for, what it was
asked, what it touched and what came back, and the log is kept where
the agent itself cannot change it. Alert rules are agreed for the
system. They can cover a burst of reads, use outside expected hours
or access beyond the agent’s agreed data scope. When something does
go wrong, the record is what turns an incident into a timeline
rather than a guess.

A lock is tested, not trusted

We test the access controls within the engagement's scope and record
what was attempted, refused or still needs attention. Tests can
cover denied tool calls, rate limits, credential isolation and
prompt-injection attempts. The result is a reviewable verification
record and a set of repeatable checks, with their limits stated.
Incident work and credential rotation are scoped separately when an
existing exposure needs containment.

The work behind it

We have done this work on servers, identity systems and platforms.
We wrote 1,000 RBAC rules across Linux and database servers, led
access management across M365 and single sign-on, and ran a
continuous security scanning programme with 21 written audits. We
split every operational secret per host, added a one-second
preflight for secrets, found three locks on the door, none of them
locked, and proved the intrusion ban works on every run. We built
fail-closed rate limiting and wrote a scope rule after a leak. How
we keep our own house is on how we work securely.

How an engagement runs

It starts read-only. We map the agents in the agreed scope, their
credentials and what they reach, and hand you the findings in order
of risk, each with its reason and the fix. The fixes are then made
in your configuration and your repository, through your review, with
checks that fail when covered access paths exceed the agreed
permissions. You keep the map, the record and the tests — the file
you hand to an auditor or an insurer — and the data agents read can
be permissioned the same way. It is bought like the rest of our
work, on the terms in what it costs.

If agents already hold keys to your systems, tell us which ones
(read by the engineer).

hThis page on the web	URL:https://data.engineer.company/ai-security/	data.engineer.company	70
