Ten simulated bots walked past 65.3% of the popular sites DataDome tested

The test set covered simple scripts, bots presenting themselves as recognized AI services, tools with forged browser fingerprints, and automated instances of real browsers, and the requests were sent through residential internet addresses in the United States, Canada and France ^[22]. Subtracting the sites that challenged everything and the sites that challenged nothing leaves 32.3% with partial coverage ^[3].

On growth, 124% against 13.2% is roughly 9.4 times the human rate ^[1]. The 82.3% rise in AI agent and crawler traffic is about 6.2 times it ^[2]. Human traffic's share of requests fell over the period, and bad bots accounted for most of that change ^[7].

Scraping was the dominant activity, at 70.9% of bad bot traffic ^[10]. In a 30-day sample, simple bots were the largest single share, cheap to run and still effective against sites with limited protection, used for scanning, reconnaissance and attacks on easier targets ^[12]. Average-complexity bots built on browser automation frameworks turn up in scraping, fake-account creation and credential testing ^[13]. Advanced operators keep their better tooling separate from the basic bots so that a simple automation signal does not expose a campaign ^[14]. Credential stuffing was flat year on year, with attackers running high-volume attacks, withdrawing to refresh credential lists and rotate infrastructure, then returning at full scale ^[9][11].

Browser-side signals do not cover this. They can be suppressed or imitated, and requests aimed at APIs, mobile endpoints or server-side services may never execute client-side code at all ^[15]. User-agent allowlists fail the other way: malicious software can copy the string of a recognized AI service and arrive as a trusted crawler, while an unfamiliar bot misses the rules written for known services ^[20]. That pushes detection onto the sequence, speed, volume and context of each interaction, because one request may not reveal a visitor's purpose ^[19].

"Organizations are being asked to make more nuanced decisions with defenses that are still largely built around binary choices," said Jerome Segura, VP of Threat Research at DataDome ^[17]. "The ability to distinguish a legitimate AI assistant from a credential-testing bot or an automated account-abuse campaign will be critical to protecting customers without blocking beneficial activity," he said ^[18].

The two halves of the report measure different populations. Traffic share and growth come from trillions of requests to DataDome's own customer sites ^[5]. The failure rate comes from tests against popular websites, and the report does not state how many sites were in that set ^[25]. Volume gave little indication of protection: high-traffic sites performed at a similar level to sites with smaller audiences ^[23]. DataDome attributes the gaps at large organizations to complex infrastructure, disconnected tools and legacy systems ^[24].