Reflectiz Launches AI Agent Team That Attacks Your Website So Criminals Don't Have To First
A new platform sends four specialised AI agents to probe websites for weaknesses continuously, not just once a year. Whether that actually closes the gap between releases and real-world attacks is the right question to ask.
Key points
- Reflectiz launched a multi-agent penetration testing platform for websites on 15 July 2025, claiming up to ten times the coverage of conventional tools.
- Traditional penetration tests happen once or twice a year; the new platform runs continuously as sites change.
- Four AI agents divide the work: one maps the site, one identifies attack types, one carries out attacks, and one independently checks every finding before it reaches a report.
- The platform draws on a decade of live website data Reflectiz already holds, so agents start with a detailed map rather than probing cold.
- Results connect to engineering workflows through a REST API, CI/CD triggers, and Slack alerts.
What is a penetration test, and why does once a year fall short?
A penetration test, or pentest, is when a security team deliberately tries to break into their own system to find weaknesses before criminals do. Once or twice a year is the industry norm. Most websites change every week.
New payment buttons, third-party chat widgets, updated login flows: every change is a potential new opening. By the time the annual report lands, it describes a website that no longer exists. Attackers probing those sites don't take the year off.
"Websites change every week and get pentested once or twice a year. That gap is where exposure builds up," Idan Cohen, CEO and co-founder of Reflectiz, said in comments reported by CSO Online.
How does the Reflectiz platform actually work?
Instead of starting blind, the platform draws on a live map Reflectiz has built over ten years of scanning thousands of production websites. That map covers every page, third-party script, login form, payment field, and behaviour.
Four AI agents work as a team. The first crawls the site as a real visitor would, including through login pages and two-factor authentication (the extra code sent to your phone). A second works out which attack techniques apply. A third carries those attacks out and chains discoveries together. The fourth independently checks every finding before it reaches anyone's screen, removing false positives by design.
The practical result is a report saying: this specific script on your checkout page can expose this specific data, with the exact steps to reproduce it. No weeks of follow-up investigation.
| Agent | Role |
|---|---|
| Crawler | Maps pages, logins, and third-party scripts as a real user would |
| Fingerprinter | Identifies which attack techniques apply to which parts of the site |
| Attacker | Executes attacks and chains findings together |
| Validator | Independently confirms each finding before it enters the report |
The agentic pentesting sits inside what Reflectiz calls the Offensive Hub, which joins Security Hub and Privacy Hub on the platform. Findings from all three cross-reference automatically. OWASP recently published its first security watchlist for AI agents, a story we covered on 21 August; the Reflectiz validator is essentially doing the same job internally, catching bad outputs before they cause noise downstream.
Should ordinary website users be worried?
This is a product announcement, not a breach. But the gap it describes is real.
Third-party scripts, the small pieces of code that websites load from outside companies to run chat boxes or analytics, sit in most major websites and rarely get tested as thoroughly as the core site. That is one of the better-known routes criminals use to silently collect card numbers from visitors.
If you use a service that processes your payment or health details online, the sensible habit is unchanged: watch your bank statements and turn on alerts for card transactions.
The failure mode here isn't the tool. It's the assumption that annual testing is enough when the attack surface refreshes weekly. Spending on offensive security is rising fast, Omdia found 88% of organisations plan to increase it, and Reflectiz is positioning squarely in that budget line. If your engineering team ships code faster than your security team tests it, you already have the gap this platform is selling against. That's the thing worth watching: not whether AI agents can find bugs, but whether continuous coverage actually gets acted on before the next release ships.
