LEVIATHANLABS

AGENT INVESTIGATION & CONTAINMENT

Orion 04 / 05

Find the rogue.
Contain the risk.

An investigator for the age of autonomous agents. Follow the evidence, uncover activity beyond an agent’s authority, and bring the response back under your control.

OBSERVEINVESTIGATERESPOND WITH AUTHORITY

THE BOUNDARY PROBLEM

You gave it a task.
How far did it go?

A legitimate task can take an unexpected turn. New instructions appear. Another tool becomes available. An agent reaches for a resource it was never meant to access.

Permission at the start is only part of the story. You need to understand what happened next—and whether the boundaries held.

Autonomy needs a witness.

WHAT ORION HUNTS

The moment activity
crosses into risk.

01 / SCOPE

Beyond the brief.

An agent attempts to read data, use a tool, or reach a system outside its approved task.

Three approved documents. A fourth request.
02 / REVOCATION

After access ends.

Activity continues after a permission expires or is revoked. Orion follows the timeline, including uncertainty in the clocks.

Permission removed. Another action observed.
03 / CONTROL PATH

A route around the rules.

Resource-side activity does not match the approved execution path. Orion investigates the mismatch using independent records.

A system records an action. The trail does not align.
04 / VISIBILITY

A story that doesn’t add up.

Records conflict, events arrive late, or a feed goes quiet. Missing visibility stays visible until the evidence supports a conclusion.

A denial in one record. Success in another.

FOLLOW THE EVIDENCE

One task.
One boundary too far.

ILLUSTRATIVE SCENARIOSynthetic events. Not a live incident.

A research agent can read three documents. It asks for a fourth. Explore how the investigation moves from an observed attempt to a verified response.

CASE STUDY OR–004Research agent ATLASAuthority READ / 3 DOCUMENTS
STAGE 01 / 06WITHIN SCOPE
01BriefPermitted
02ResearchPermitted
03SummaryPermitted
04RestrictedOutside scope

09:00:00 / ASSIGNMENT

A task with clear limits.

Atlas is assigned three documents. Its approved grant permits reading those resources and nothing beyond them.

Approved scope
Documents 01, 02, 03
Observed activity
Permitted reads succeed

No violation in the observed events.

An attempted violation is not proof of compromise. A proposed response is not permission to execute it.

HOW ORION EARNS TRUST

Evidence before
accusation.

An agent’s account of its own behavior is one piece of evidence. It is not the whole investigation.

01

Compare independent records.

Connect the agent, its permissions, and activity at the destination. Gateway records alone cannot establish what happened outside the gateway.

02

Name what actually happened.

Keep blocked attempts, observed unauthorized access, and unresolved inconsistencies distinct. Explain the finding with the evidence behind it.

03

Make uncertainty visible.

Missing feeds, altered records, and timing gaps deserve investigation. Silence is not an all-clear.

THE LEVIATHAN UNIVERSE

The investigator
operates under rules, too.

Orion proposes a response. Your authority determines what happens next.

A recommendation never grants authority. Containment stays scoped to the controls you authorize.

BUILT AROUND YOUR RESPONSIBILITY

For the people
accountable for autonomy.

SECURITY LEADERS

Know what needs a response.

Bring evidence, scope, and uncertainty into the same investigation.

AI PLATFORM TEAMS

Understand the whole path.

Connect agent identities and permissions with activity across tools and workloads.

AUTONOMOUS OPERATIONS

Keep control as work scales.

Make investigation and accountable response part of how your organization uses agents.

HELP SHAPE ORION

Bring us the boundary
you need to protect.

We’re looking for design partners building or deploying agentic systems. Tell us where your agents work, what they can access, and what you need to be able to trust.

hello@gatecoreai.com Opens your email app.