Skip to main content
Autonomy · Persistence

What autonomous persistence means under authorisation

A red team holds a foothold so that access can be shown to be durable rather than momentary. When the thing holding it is a platform rather than a person, one question decides whether that is acceptable, and it is not a technical one: under whose authority, and bounded by what. B-52 answers it with a document signed before the run rather than a judgement made during it.

Yash K · · Autonomy and its limits · about 10 min

The word itself

Persistence is a foothold held across time

It is the part of an offensive engagement that proves access was durable rather than momentary, and it is the part a customer has the most right to ask questions about.

A foothold that lasts as long as one session proves that a door opened. A foothold that survives a reboot, a session expiry or a credential rotation proves something harder and more useful — that somebody who got in on a Tuesday would still be in on the Friday, and that nothing around that host closed the gap in between. Red teamers call the second one persistence, and it is worked deliberately rather than stumbled into: an implant, a scheduled task, a stored credential, a token that outlives the login that issued it. It is also the single action in an offensive engagement that most obviously belongs to the customer to authorise, because unlike a scan or an exploit proof it leaves something behind and keeps it there. On a B-52 engagement it therefore sits behind a written approval rather than inside the scope, and the rest of this article is about why those are two different documents.

The same moment, twice

Where an operator would take over, and what happens instead

Reaching a foothold is one problem. Deciding what to leave behind, which host to move to next and when a chain has proved enough is a different one, and it is the half a conventional engagement hands to an operator.

01 With an operator

A conventional engagement

The foothold
A tester reaches a shell, a session or a working set of credentials on the first host inside the perimeter.
The decision
A human operator takes the keyboard. What to install, how long to hold it and how far to move are judged in the moment, against that operator’s reading of the situation and of the scope in front of them.
The record
The judgement is reconstructed afterwards from notes and from the report. The authority for it was the engagement letter, read in the moment by the person acting on it.
02 With nobody in it

A fully autonomous B-52 run

The foothold
The same. Discovery, testing and chaining run against the authorised scope without checking in, and a foothold is reached in the course of that work.
The decision
There is no operator to hand it to. B-52 establishes persistence itself, inside the boundary a written approval has already named — and where no such approval exists, it stops there, proves what it reached and reports.
The record
The authority is a document signed before the run, naming what was permitted and carrying the date it was given. Nothing is reconstructed afterwards, because nothing was judged at run time.

Settled in advance

What the authorisation fixes before anything runs

Each of these is a judgement on a conventional engagement and a sentence on this one. None of them is decided by the platform while the run is going.

The questionWhat is written down, and when
Under whose authority A named person on your side signs the scope before the engagement begins. In the fully autonomous model there is no operator behind that signature to widen it later, so it is not a formality that precedes the real permissions — it is the permission.
Bounded by what The hosts or the segment inside which persistence and movement past the entry host are permitted, named in their own written approval rather than inferred from the scope. Approving movement inside a named segment authorises it inside that segment; reaching past it is a new decision.
The boundary is a place, written down in advance — not a rule the platform applies to a situation it meets.
For how long For the engagement the approval was given for. It is not a standing permission, it does not travel to the next piece of work, and a second engagement asks again.
What it does not carry with it Live credentials and real customer data sit behind their own gate. So do destructive and state-changing actions against a production system. Holding a foothold releases neither of them, and neither is implied by an approval to hold one.

Under whose authority

What is written down, and when
A named person on your side signs the scope before the engagement begins. In the fully autonomous model there is no operator behind that signature to widen it later, so it is not a formality that precedes the real permissions — it is the permission.

Bounded by what

What is written down, and when
The hosts or the segment inside which persistence and movement past the entry host are permitted, named in their own written approval rather than inferred from the scope. Approving movement inside a named segment authorises it inside that segment; reaching past it is a new decision.

The boundary is a place, written down in advance — not a rule the platform applies to a situation it meets.

For how long

What is written down, and when
For the engagement the approval was given for. It is not a standing permission, it does not travel to the next piece of work, and a second engagement asks again.

What it does not carry with it

What is written down, and when
Live credentials and real customer data sit behind their own gate. So do destructive and state-changing actions against a production system. Holding a foothold releases neither of them, and neither is implied by an approval to hold one.

The order of authority

Four steps, and the human ones are both at the front

Order is most of the argument here. Each step is spent before the next becomes possible, and the steps a person takes come before the run rather than during it.

The whole argument

Agreed before it started, rather than judged while it ran

One condition decides whether an autonomous system holding access is acceptable at all, and every document above exists to meet it.

An autonomous system holding access inside somebody’s estate is acceptable on one condition: that its limits were settled in advance by the person who owns that estate. So they are settled in advance. The scope authorisation names the targets and the boundary drawn around them. A separate written approval names the hosts or the segment inside which persistence and movement are permitted, and carries the date it was given. Anything past that boundary is a fresh decision rather than an inference the platform is entitled to make on its own. A conventional engagement can afford to leave some of this to be worked out during the run, because there is an operator on the far end who will work it out — in the moment, against their own reading of the scope in front of them. Autonomy removes that reading, and what replaces it has to exist on paper, and has to exist first. That is the trade this model makes, and it is why the scope document carries more weight here than in either of the other two.

Inside and outside

What an approval for persistence covers, and where it stops

An approval is given for a boundary rather than for a single action inside it. That cuts both ways, and both directions are set out here.

What an approval for persistence covers, and where it stops
StateWhat it meansWhat follows
Inside the named boundary B-52 establishes and holds persistence itself on the hosts or the segment the approval names, for the engagement the approval was given for. Proceeds without asking again.
A further host inside that boundary Movement to another host the approval already covers. B-52 does not return for a second signature on each host it reaches inside a boundary you have already drawn. Proceeds without asking again.
Past the named boundary A host, a segment or an environment the approval does not name — whether or not it looks equivalent to one that is named. Resemblance is not authorisation. Stops, and asks in writing.
Live credentials or real customer data Using a real account or reading a real person’s records once a path to them has been proved. It is a separate gate with a separate approval, and holding a foothold does not release it. Stops, and asks in writing.
No approval was given at all Terminal The run does what the scope authorised, reaches the foothold, proves it with the request, the response and the steps that reproduce it, and reports. A gate that was never released is a gate that was never released. No persistence is established.
Key
  • Authorised in advance, in writing, for a named boundary
  • Stops and waits for a further written approval
  • Nothing was released, and the run reports what it reached
  • TerminalNo state follows this one

The turnaround figure

How long a run with nobody in it takes

One turnaround figure is published for B-52. It belongs to the model this article is written about, and this is the measurement behind it.

Turnaround on the fully autonomous model As of 2026-09-15
  • A median of one to three business days from scope sign-off to report delivery, observed across engagements run on the fully autonomous model.
  • It is a median taken from real runs rather than a service level, so individual engagements land on either side of it.
  • Where the movement gate is released at scoping rather than mid-run, the run is not waiting on a signature while it is going.

Deliberately excluded

  • The expert-verified and human-led models. No turnaround figure exists for either of them, so none is stated.
  • Physical, hardware and wireless testing, which are out of scope for B-52 in every coverage class.

Persistence is authorised in writing, before the run starts

A scoping call settles the scope document and the movement boundary in the same conversation. One scan of one application or one target is $500, and a paid trial is $299.