PTaaS vs autonomous pentesting: who leads the test, and what comes back

PTaaS and autonomous pentesting are not two versions of the same product. PTaaS is a way to buy human-led testing through a platform. Autonomous pentesting is a way to run part of a test in software. Most programmes that can afford both should use both, with a written line between them.

What PTaaS actually is

PTaaS, penetration testing as a service, is a delivery model rather than a testing technique. The providers who sell it describe it in similar terms.

Bugcrowd calls it "a delivery model that combines human-led penetration testing with a cloud platform". Its platform covers scoping, collaboration, findings, remediation tracking, reporting and retesting. Cobalt describes PTaaS as a platform-delivered model for offensive security testing. Its pentesters lead validation, exploitation and complex testing, while AI supports discovery and reporting. NetSPI describes its PTaaS as an expert manual pentesting team enhanced by automation.

Read across those pages and the common elements are clear:

• Humans do the core testing. Bugcrowd says plainly that PTaaS does not mean replacing pentesters with a scanner.
• A platform takes over the consulting workflow. You scope and launch a test there instead of through a statement of work and an email thread.
• Findings arrive during the test, not only in a final report, with remediation guidance attached.
• You can talk to the testers while the engagement runs.
• Retesting of fixes is requested through the same platform.
• Results flow into existing tools, such as ticketing systems.

So the honest comparison is not "humans vs software". It is "human-led testing bought through a platform" vs "software-led testing supervised by your own people".

The line is already blurring

Several PTaaS providers now add AI to the human-led model. Synack describes itself as an AI and human powered pentesting platform, and includes an agent it calls Sara, the Synack Autonomous Red Agent. Cobalt has announced AI features for scoping, reconnaissance and reporting. It has also announced an Autonomous Pentest product. Cobalt says that product is designed to complement, not replace, human-led pentesting for compliance-driven engagements.

That is a sensible direction, and it changes the buying question. Do not ask whether there is a machine involved. Ask three things instead. Who leads the test on each asset? What is the machine allowed to do without a person? What proof comes back with each finding?

What we mean by autonomous pentesting

Arxiis is a human-supervised, autonomous offensive-security platform. A crew of LLM operator-agents runs an authorized web-application or API test end to end. It will not engage a non-local target without a written SCOPE + ROE authorization file. A scope guard checks every request against that file, and a human operator approves each escalation.

Every command the crew runs is written to a tamper-evident, hash-chained execution ledger. A finding is labelled proven only when its evidence binds to a real ledger record. Anything else is labelled needs review.

Our measured results on XBOW-104 (19 October 2026) are 56% raw and 44% confirmed across all levels. On common web bugs we solve about 82%, level-1 tier only. On hard chained and blind attacks we solve 0–42%, a known weak spot. Our measured false-positive rate across XBOW-104 sweeps is 0%.

The lead belongs to your operator, not to an outside testing team. That is the real difference from PTaaS.

Where human-led PTaaS still wins

Logic flaws and chained exploits. Synack's own page says its researchers find what automated tools miss, naming logic flaws and chained exploits. Our numbers agree on the second point: 0–42% on hard chained and blind attacks is a known weak spot for Arxiis. We have not measured business-logic testing at all. If that is where your risk sits, human testers should lead.

Breadth of target types. PTaaS providers describe testing well beyond web applications. Cobalt lists network, cloud and red team work alongside web apps and APIs. NetSPI lists network, cloud, social engineering and more. Arxiis is measured on web-application and API classes only. For anything else, it is the wrong tool.

A tester you can talk to. PTaaS providers stress live collaboration between your team and the testers. A person can explain a finding, adjust focus when something interesting turns up, and answer your engineers' questions. Our operator console serves your operator, not your developers.

Engagements that expect a human-led test. If an auditor, customer contract or internal policy expects a human-led penetration test, buy one. Cobalt itself positions its autonomous product as a complement for compliance-driven work, not a substitute. Ask your auditor what they accept before you assume anything.

Judgment about business severity. Human testers can rank findings by what the business would lose. Our labels describe evidence, not business impact, and we make no breach-impact claims.

Where supervised autonomous runs fit

Between engagements. A human-led test happens at the intervals you buy. Your applications change between them. Supervised autonomous runs can test the same authorized scope after each significant change, under the same written authorization. That is the case we make for continuous testing, with every run authorized and every finding bound to evidence.

The routine sweep. Common web-bug classes such as SQLi, SSTI, SSRF, XXE, JWT and LFI are where Arxiis is strongest. Clearing them in software means your PTaaS hours go to the work only people do well.

An evidence trail with a fixed shape. Every run writes the same kind of ledger record. Each proven finding points at its record, and each unproven one says so. That gives an auditor one consistent artefact across many runs.

Hard stops. No authorization file, no engagement. The scope guard blocks hosts outside the file. When the budget or time window runs out, the watchdog stops the run and writes a partial report that says it is partial. It does not claim a completion it did not reach.

How to combine them

One authorization, two lanes. Write one SCOPE + ROE document for the programme. Name which assets get human-led PTaaS, which get supervised autonomous runs, and who approves escalation in each lane. Overlapping, undocumented scopes are how two tools end up testing the same host on the same night.

Give the people the hard part. Within your agreement with the provider, share your autonomous needs review queue with the PTaaS testers. Point them at the logic-heavy workflows and the chained and blind classes we are weak on. Senior hours should not be spent re-finding routine bugs.

Hold both lanes to one evidence standard. Bugcrowd says each validated issue includes evidence. Ask your provider what that artefact is, and whether it includes the request and response behind the finding. Ask the same of any autonomous tool, and require proven or needs review on every finding.

Line up the calendar. Run the autonomous lane before a PTaaS engagement, so testers start with routine issues already filed. Run it again after fixes ship, and compare the new ledger records with the old ones.

Report the lanes separately. Show leadership one view, but label every finding by source and by evidence status. A human-validated finding and a proven autonomous finding are both actionable. A needs review item is not yet a finding.

A quick decision table

What to ask both kinds of vendor

Ask a PTaaS provider and an autonomous vendor the same questions, in writing.

1. Who or what leads the testing on each asset, and what may the software do without a person?
2. What artefact sits behind each reported finding, and can we inspect it?
3. What happens when a request would go outside the written scope?
4. Which target types and vulnerability classes are you not validated for?
5. How do retests work, and what record does a retest leave?

The answers will differ, and they should. What matters is that each one is specific and written down.

The trade-off

If you can afford only one and your risk is mostly in logic, buy human-led PTaaS. We have not measured Arxiis on logic flaws, and we would rather say so here than after you have bought it. If your risk is many web applications changing every week, supervised autonomous runs earn their place. Keep a human-led test at intervals all the same. The open question is how far each side moves towards the other. And how honestly each one labels what its software did alone.

Frequently asked questions

What is PTaaS?

Penetration testing as a service is a delivery model for human-led penetration testing. Providers describe a cloud platform where you scope and launch tests and see findings during the engagement. You can also talk to the testers, request retests of fixes and send results into existing tools.

Is PTaaS the same as autonomous pentesting?

No. In PTaaS, human testers from the provider lead the test and a platform manages the workflow. In autonomous pentesting, software runs much of the test. In Arxiis, that software is a crew of LLM operator-agents supervised by your own operator under a written authorization file. Some PTaaS providers now add AI, so ask who leads each test.

When should I choose human-led PTaaS over autonomous pentesting?

Choose it when the risk sits in business logic or chained attacks, or the scope includes network, cloud, social engineering or red team work. Choose it when an auditor or contract expects a human-led test. And choose it when you can afford only one and your risk is mostly in logic.

Can I use PTaaS and autonomous pentesting together?

Yes. Write one authorization that names which assets each lane covers and who approves escalation. Run supervised autonomous tests between PTaaS engagements. Hand the needs review queue and the hard chained classes to the human testers. Hold both lanes to the same evidence standard.

Does Arxiis replace a PTaaS provider?

No. Arxiis covers web-application and API classes and solves 0 to 42% of hard chained and blind attacks on XBOW-104, a known weak spot. It fits between human-led engagements, not in place of them.