
On 7 April 2026, Picus Security published a sponsored article on BleepingComputer. It argues that automated pentesting tools hit a "PoC cliff": they surface plenty of findings on the first run, then little new by about the fourth or fifth. The piece is vendor marketing for breach and attack simulation (BAS), and its numbers come without methodology. Still, the underlying engineering point is sound, and it matters to any team that treats one green scan as proof it is secure.
What happened
The article, by Picus security research engineer Sila Ozeren Hacioglu, makes several claims:
- The PoC cliff. Automated pentest tools follow chained attack paths. Once you fix the path a tool prefers, it gets blocked early and never exercises the later steps, so new findings drop off quickly with each run.
- The "validation gap". This is the difference between what an organisation has actually tested and what it reports as validated.
- Pentesting and BAS answer different questions. Pentesting asks whether an attacker can get from A to B. BAS runs many independent, atomic tests to check whether controls such as EDR, WAF and SIEM rules actually detect or block MITRE ATT&CK techniques.
- Six validation surfaces. Network and endpoint controls, detection and response, infrastructure and application paths, identity and privilege, cloud and containers, and AI systems. The article claims automated pentesting alone fully validates none of them.
It also says over 60% of findings are wrongly classified as high or critical and about 10% are genuinely exploitable. These are vendor figures with no published method.
Who is affected
- Security and platform teams that run automated pentesting (or "autonomous red teaming") products on a schedule and report the results as coverage.
- Engineering leaders who rely on those reports for compliance or risk sign-off.
- Application teams whose code paths a network-centric pentest tool rarely touches, such as business logic, authorisation and multi-tenant isolation.
What it means for developers and engineering teams
Strip away the vendor framing and three useful ideas remain.
Chained tests stop where the chain breaks. If step 2 of an attack path is blocked, the tool never tells you whether steps 3 to 6 would have worked, so a fix for one weakness can hide several others. Developers know this pattern from integration tests that fail fast. Atomic tests, which exercise each control or technique on its own, give you a fuller picture.
Diminishing findings don't mean you are secure. A deterministic tool with a fixed playbook will eventually exhaust that playbook. A quiet fifth run tells you about the tool's scope, not about your exposure. Measure coverage explicitly, for example as a matrix of techniques or surfaces tested, rather than counting findings.
Application-layer risk lives in your code. Broken object-level authorisation, IDOR, tenant isolation bugs and business-logic abuse are the most common serious web vulnerabilities, and generic network pentesting or BAS rarely finds them. Those need tests written by people who understand the app: authorisation unit tests, DAST against authenticated sessions, and periodic human-led testing.
What to do now
- Write down what each tool actually covers across network, detection, application, identity, cloud/containers and AI. Mark the gaps as accepted or to be addressed.
- Add atomic detection tests. Open-source options like Atomic Red Team let you check that individual EDR or SIEM detections fire, whatever any attack chain does.
- Put security tests into CI: SAST, dependency and container scanning, IaC policy checks and authorisation tests for sensitive endpoints, so coverage keeps up with code changes.
- Rotate the methodology. Change tools, playbooks or targets between runs, and schedule human pentests for business logic and novel paths.
- Challenge severity scores. Prioritise by exploitability in your environment (reachability, compensating controls, CISA KEV status), not by raw CVSS.
- Ask vendors the hard questions: which surfaces they cover, how they tell exploitable from theoretical, and how they de-duplicate findings from your other tools.