Your website has a new front door: Generative AI wants in, and the attackers are already there

Generative AI has created a strange new security problem: many organisations now need to welcome some non-human visitors in order to stay visible in generative search, generative engine optimisation (GEO), and emerging agent-driven customer journeys, while stopping other non-human visitors that look alarmingly similar. Public-facing websites are no longer dealing only with search crawlers and obvious bots. They are increasingly dealing with generative AI crawlers, AI search systems, user-driven assistants, and browser-based agents that can read pages, compare products, fill forms, and move towards transactions. That shifts the security question from "bot or not?" to "which automation should we trust, where, and under what conditions?".

The numbers show why this deserves urgent attention. In 2025, AI-driven traffic grew 187% across the year, while traffic from AI agents and agentic browsers surged 7,851% year on year. More than 95% of observed AI-driven traffic was concentrated in retail and e-commerce, streaming and media, and travel and hospitality. At the same time, the median share of traffic attempting scraping attacks approached 20% in 2025, post-login account compromise attempts more than quadrupled year on year, and the average organisation saw 402,000 post-login compromise attempts. Across all interactions analysed, just 0.5 percentage points separated benign from malicious automation.

This talk argues that the old playbook, built around robots.txt, blanket bot blocking, CAPTCHAs, and static allowlists, is no longer enough. The same web surfaces that benign generative AI needs to access for discovery and customer value, such as product pages, account pages, authentication flows, APIs, and checkout, are exactly where attackers are concentrating effort. I will present a practical, vendor-neutral framework for separating training crawlers, AI search, user-authorised assistants, agentic browsers, and hostile automation; for deciding what should be allowed, challenged, rate-limited, or blocked by page type and transaction stage; and for combining identity, declared purpose, behavioural evidence, step-up controls, and transaction-aware trust. This is not just a web security problem. It is now a shared design problem for security, fraud, digital, and search teams. Attendees will leave with a concrete approach to support legitimate generative AI access without handing adversaries a wider door into the business.