AI pentesting arrives in attack surface management: what changed in 2026

What was announced?

This post covers announcements made between June and September 2026, as the vendors describe them.

• On 16 June 2026 CyCognito announced Continuous AI Pentesting, which it says runs AI-driven offensive testing continuously across the external attack surface through its Target Graph architecture.
• On 24 July 2026 watchTowr introduced Project Red, an AI-based capability that reproduces emerging vulnerabilities to validate client exposure. watchTowr states it reproduced CVE-2026-63030 in 22 minutes.
• Hadrian describes agentic AI that exploits issues to prove impact, and publishes a price for NOVA, an agentic penetration test, from 1,250 euros per test excluding VAT.
• On 8 September 2026 Tenable said it is bringing Anthropic's Claude Mythos 5 into Tenable One, starting with Tenable One Adversary View for finding attack paths and remediation options.

These are four different things. CyCognito and Hadrian describe continuous offensive testing of exposed assets. watchTowr describes reproducing a specific new vulnerability fast enough to test for it the same day. Tenable describes AI analysis of attack paths and remediation options inside its exposure platform.

Why does it matter for proof?

Proof has always been the expensive step. Version matching is cheap and noisy; a human penetration test is precise and periodic. AI-driven testing aims at the middle: precise enough to show an exposure is real, cheap enough to run all the time. If it works as described, the PROVEN station on more closure tracks fills in, and the queue of unverified findings shrinks.

What does it not change?

Three things stay the same.

1. Containment. None of these announcements adds a compensating control. A proven exposure still needs a WAF rule, a playbook or a fix.
2. Ownership. Testing harder does not tell you who owns the asset.
3. The retest. Proof before the fix is half the loop. The same engine has to check again after the fix, and not every vendor describes that.

What should buyers ask?

1. What does the AI system do to a live asset? Which actions are allowed, and which are excluded by design?
2. What evidence does a proven finding carry, and can an owner reproduce it?
3. How does a human review or stop a test?
4. Does the system retest automatically after the owner reports a fix?
5. Is it priced separately? CyCognito already prices exposure validation on a different unit from ASM, and Hadrian prices NOVA per test.

Does this change any score here?

Our proof scores already reflect these pages: CyCognito, watchTowr and Hadrian score 9 on proving exploitability, and Tenable ASM scores 5 because the ASM product itself does not document exploit testing. We will revisit when vendors publish more detail on how the new systems behave in production.