The pitch lands about once a quarter now. An AI agent that tests your systems like a senior pentester, priced on a subscription basis, with results in hours instead of weeks. If you run technology or security for a bank, a fintech, or a trading platform, you have probably sat through one of these demos this year. The technology is real. The question is what you would actually be buying.
ClarenSec uses AI-assisted tooling on live engagements. It has improved reconnaissance, coverage and triage speed, but it has not taken responsibility for a finding, understood a client's business environment, or been used to make a judgement call on a production system. Those boundaries matter more than the sales label on the tool.
What changed by 2026
The old automated vs manual penetration testing argument was easy, because the automation was dumb: a scanner matched version strings and signatures, and a human did everything else. Some current AI-assisted scanners and agents can read responses in context, pattern-match known bug classes across large codebases, adapt payloads instead of firing a fixed list and group duplicates before a person sees them. Some can also draft a plausible finding writeup. Capability varies by tool and target.
What has not changed is the category of question the tool answers. It is still asking "does this look like a problem someone has seen before?", just at far greater speed and scale. That is not the same job as an attacker deciding what your systems can be made to do.
What automation now does well
Four things stand out:
- Asset discovery: mapping the environment, the forgotten subdomains, the staging box someone exposed in 2023.
- Breadth: reachable endpoints and parameters get a consistent first pass.
- Regression checking between tests: confirming that the last engagement's fixes stayed fixed, instead of waiting six months to find out.
- Known-CVE matching: fingerprinting versions against published vulnerabilities.
What it still cannot do reliably
No single observation covers every tool. In our engagements, the clearest gaps are business logic abuse and authorization flaws that depend on roles and workflow context.
Multi-step attack paths are another gap. The findings that move money are often not one issue; they are an information leak feeding a password-reset weakness feeding an internal service that trusted the wrong header. Chaining takes a hypothesis about a specific environment, pursued across steps that each look harmless alone. Judging real-world impact also depends on knowing what the affected system does and who can reach it.
There is also the old problem in new clothes. Automated scanners produce false positives that need human validation, and an AI layer that drafts confident, fluent prose around a false positive makes it more dangerous. We have written before about the difference between a vulnerability assessment and a penetration test. In our work, AI-assisted tooling has improved coverage and triage during the assessment without replacing manual exploitation.
Proof of concept needs a person accountable for it
A finding without proof is an opinion. This is why every finding in a ClarenSec report is verified manually with proof of concept before it ships, a rule that sits at the centre of our testing methodology and reflects the exploitation phase in PTES and the attack phase in NIST SP 800-115. The record that should not have been reachable, the request that worked, the demonstrated chain: producing that evidence responsibly against production systems is a judgement call, made in real time, by someone who can stop.
For PCI DSS work, Requirement 11.4 calls for penetration testing by a qualified internal resource or qualified external third party, with organizational independence. It does not require a QSA to perform the test. ClarenSec's own reporting standard goes further: every reported finding is manually verified and carries evidence the client can reproduce.
How the work divides
On our engagements, AI-assisted tools help map the testing environment fast, check the known-vulnerability databases and correlate results across a large scope. This gives the senior penetration testers more useful leads and more time for the work that requires context: understanding business context, following promising findings across systems, proving impact and setting severity against the role of the affected system. AI broadens the analysis. Our testers turn useful leads into verified findings and remain accountable for the final report. That is how we run a modern penetration testing engagement in 2026.
- AI-assisted automation has improved reconnaissance, coverage, regression checking and triage speed.
- Business logic abuse, chained attack paths, authorization flaws and impact judgement still require a tester who understands the environment.
- PCI DSS requires a qualified, organizationally independent testing resource. ClarenSec also requires manual verification and reproducible evidence for every reported finding.
