Resemble AI launches RAD-S1 to spot AI agents that pose as humans

Resemble AI has released RAD-S1, a behavioral model that tries to tell whether a website visitor is a person or an AI agent, even when the agent is disguised. According to the company's internal benchmark, it caught 98.8% or more of agent visits that hide their automation, while six other bot detection tools caught between 3.1% and 14.8%.
- Resemble AI released RAD-S1, a behavioral model that classifies visitors as people, computer-use agents, browser automation or scripted clients.
- In Resemble's internal test of more than 2,000 agent visits, RAD-S1 identified 98.8% or more of hidden agents, versus 3.1% to 14.8% for six other tools.
- On agents that do not hide, all seven tools scored well, with RAD-S1 at 98.7% or higher.
- At the default Balanced setting, RAD-S1 flags around 5 to 6% of human visits as agents, and the benchmark covered agent visits only.
- Resemble recommends using the verdict as one signal alongside existing bot detection.
Why agents are hard to classify
Resemble AI says most sites still struggle to know when a visitor is an AI agent. Tools such as Muse, Claude in Chrome and ChatGPT agents run in real browsers, often inside a customer's signed-in session. Attackers can use similar tools with stealth browsers that make activity look like it comes from a person.
The post points to businesses already dealing with the uncertainty. It quotes Charles Schwab CEO Rick Wurster on why the bank stops outside AI agents from moving client money dozens of times a week: "We can't control it, the client can't control it, something could go wrong and who's liable?" According to Resemble, JPMorgan Chase also restricts outside agents from acting inside its systems, and Amazon has blocked Meta's Muse, saying the agent did not identify itself as AI while browsing.
Resemble frames this as three questions for security teams: whether the visitor is a person or an agent, where it should go, and how to stop bad ones without slowing customers. The company argues the first question must be answered before the other two, and that it is the one existing bot detection struggles with when an agent looks human.
Declared, undeclared and disguised visitors
Resemble separates visitors by what they do and how they present themselves. A bot follows a script, such as crawling content or filling out a form at scale. An AI agent works toward a goal, such as booking a flight, and decides its own steps. On presentation, the company describes three categories:
- Declared: the visitor identifies itself in its user agent or a signed header.
- Undeclared: it does not identify itself but leaves the usual fingerprints of browser automation tools.
- Disguised: it changes its identity or browser fingerprint to look like a regular visitor.
Resemble says traditional bot detection tends to be strongest on undeclared traffic. Disguised sessions are the category RAD-S1 targets, and an agent may use a stealth setup without the person who sent it even knowing.
How RAD-S1 works
RAD-S1 combines a behavioral model with a set of rules. The browser only collects signals, and all scoring runs on Resemble's servers. A script of about 25 KB reads pointer movement, click patterns, typing flow and scrolling, plus browser and environment signals such as request headers, automation flags and Web Bot Auth signatures from agents that identify themselves. According to Resemble, keys are logged by type and never the key itself, and the script never collects typed text or passwords.
The script sends a small batch every 1.2 seconds. The first check runs after about 5 actions or 50 pointer moves, and each server check takes about 120 ms at the median and 200 ms at the 95th percentile. The signals combine into a probability that a person is present, plus a visitor class, and the verdict is final for the rest of the visit once two reads agree.
Training data includes sessions from Claude, Codex, Grok, ChatGPT, OpenAI Dot, Browser Use, Playwright and Muse, among others. The model is versioned and updated as new agents and stealth tools appear, with no fixed retraining schedule.
Benchmark results
Resemble compared RAD-S1 with Cloudflare Turnstile, BotD, F5, Google reCAPTCHA, Akamai and Vercel BotID. The test used more than 2,000 agent visits across 12 website stacks, including Vue, React, Next.js and plain HTML. It covered commercial agents with humanized mouse movement and typing, open-source automation such as Browser Use and Playwright, stealth browsers (Camoufox, Patchright and Stealth Playwright), and session takeover, where a bot takes control partway through a visit.
Against agents that do not hide, all seven tools did well. Detection ranged from 94.8% for Vercel BotID to 99.3% or more for BotD, with RAD-S1 at 98.7% or higher.
Against agents that hide, the results diverged. RAD-S1 identified 98.8% or more of those visits, while the other six tools identified between 3.1% and 14.8%. On session takeover, BotD caught about half of visits at 50.2%, and the other five tools caught fewer than 15%.
Resemble attributes the gap to inputs. Most bot detection relies on network and browser data, which stealth tools are built to fake. RAD-S1 studies behavior instead. The signals are documented, but the company says how the model weighs them is not public.
Limitations and what it means for security teams
Resemble is open about the limits. Every visit in the benchmark came from an agent, so it shows how much agent traffic each tool catches, not how often it wrongly flags people. At the default Balanced setting, RAD-S1 currently flags around 5 to 6% of human visits as agents. The company says sites will be able to adjust sensitivity from the dashboard, and that a new benchmark aimed at lowering false positive and unclassified rates is in progress. Declared agents were not benchmarked because they name themselves. RAD-S1 works on websites and web apps in a browser, and native mobile apps are coming soon.
Resemble also notes that behavior can be imitated, and some agents already try, including ChatGPT, Claude and Muse in its test. It therefore recommends treating a verdict as one strong signal among other controls and labeling traffic before deciding what to block. Its advice to security teams is to keep existing bot detection, which handles crawlers and scripts that announce themselves well, and add a layer for agents that hide.
