It runs the engagement. Not a scan.
Autonomous penetration testing means the software carries out the actions a tester would carry out, rather than producing a list for someone else to act on. That distinction is the product.
What does the agent do without a human present?
It works the engagement in phases and adapts to what it finds, in the same order a tester would.
- Reconnaissance
- Enumerates hosts, services, endpoints, parameters and technologies, and follows what it discovers rather than a fixed list.
- Authentication
- Uses supplied credentials, and where sign-up is open it registers a marked test account so it can test one identity against another.
- The hunt
- Works the vulnerability classes against the surface it mapped, choosing depth by what the target actually exposes.
- Proof
- Re-issues the request that demonstrated the issue and records the exchange.
- Writing
- Produces the finding text, the impact, the remediation and the report.
Widen its own scope. It cannot add a target, and it cannot reach one that was not authorized.
Enable a class of action. Exploitation and destructive classes are a human decision, recorded against a name.
Act on instructions found in a target. Everything a scan returns is treated as untrusted data, not as direction.
Continue past a breaker. A ceiling breach halts the run rather than slowing it down.
Where a human is still better: business logic that only makes sense if you know the company. An autonomous test will not understand that an approval step exists because of a regulator. Buy human hours for that, and use this for everything else.
Questions people ask about the autonomous part
Can the agent decide to test something outside my scope?
No. The agent proposes actions and a separate gate executes them, and that gate re-checks the specific target against your scope every time. An agent that reasons its way to a bad idea still cannot carry it out.
What happens when a page tells the agent to do something?
Nothing. All content returned by a target is treated as untrusted data. It cannot expand scope, change the action policy, authorize an action or cause a secret to be revealed.
Does it need an agent installed on my systems?
No. Testing is performed against your systems the way an external attacker would reach them, so there is nothing to deploy inside your estate.
See it work on something you own.
Set the envelope, then read what it did.