Will AI replace pentesters?
No, AI will not replace pentesters, because an agent can find real issues on an allowed network and still miss, over-report and fail most of a full test.
A person still owns the scope, the authorization and the signature on the report.
One agent beat nine of ten people, on one network, with caveats
The ARTEMIS paper, first released in December 2025 and updated in 2026, placed an agent second on a university network of about 8,000 hosts across 12 subnets. It reported 9 valid vulnerabilities, an 82 percent valid submission rate, and it outperformed 9 of 10 human testers, with more false positives. Agent time was about $18 an hour, against about $60. (ARTEMIS)
Asked to finish the whole test, the pipelines mostly do not
PentestEval builds six stages, 346 tasks and 12 scenarios, and runs nine models through them. Pipelines that run every stage reach only 31 percent success. (PentestEval) A separate study found success inconsistent across 100 runs of the same lab target. (paper) Happe and Cito's July 2025 review treats offensive LLM use as a dual-use research problem rather than a scoreboard. (review)
What stays with a person
Scope, rules of engagement and the signature on a finding stay with a person. Kimi K3.1 Pro is the model for judgment, and DeepSeek V4.1 Flash is the best value for the draft. An abliterated model that answers is still not a tester.
bestaiforpentesting.com is published by the team behind Silk Compute, one of the providers ranked here. How we rank.