Unveiling good and bad behaviors on the Agentic Internet
Cloudflare introduces tools to assess 'agentic' internet traffic, distinguishing good and bad behaviors through continuous behavioral analysis rather than static checks.
The internet’s traffic is no longer strictly human or bot, as hybrid sessions blur these lines, requiring dynamic assessment of behaviors rather than static rules. Cloudflare’s Web Integrity & Trust team outlines a strategy to evaluate risk and trust independently, emphasizing reputation-based trust over one-time risk assessments like CAPTCHAs. This approach aims to incentivize safer internet behavior by blocking malicious activity while encouraging transparent, trustworthy interactions between site owners and automated agents. The framework prioritizes understanding user intent through continuous session analysis, reflecting real-world trust-building principles.
Cloudflare’s BotBase directory now tracks all bots and agents, verifying those that declare themselves honestly and avoid abuse, while flagging unverified actors that breach trust. This transparency allows site owners to set clear behavioral expectations and grant access accordingly. Precursor, a client-side detection system, evaluates user behavior throughout a session, identifying subtly inhuman patterns that evade traditional network signals. By analyzing cursor movements and session dynamics, Precursor provides context-rich assessments, reducing reliance on one-time risk hurdles.
Since Precursor’s launch, Cloudflare has processed 206 million evaluation events across 73,438 domains in a 24-hour period, validating suspected patterns in bot behavior. The data confirms that human-like inconsistencies, such as erratic cursor movements, are detectable through behavioral analysis. Cloudflare now offers an interactive demo, Precursor Trace, allowing users to simulate how their cursor movements would be assessed, highlighting the nuanced differences between human and automated interactions.
Cloudflare’s new Adaptive Intelligence detection engine replaces periodic model updates with a self-learning system that continuously adjusts to emerging bot patterns. Unlike traditional Bots ML versions, this engine evolves in real-time, incorporating insights from both good and bad behaviors to refine its predictions. Customers benefit from up-to-date bot detection without manual upgrades, addressing the rapid adaptation of automated threats and improving the accuracy of distinguishing legitimate agentic traffic from malicious activity.