Free-tier promotion attracts a scripted assault

When AI research lab Nous Research opened its Hermes LLM to free inference for one week, the intent was accessibility. The reality was an automated siege: scripts generated datagen prompts and spread the load across thousands of fake accounts to bypass rate limits. Bulk signups slipped past the existing Cloudflare Turnstile captcha, burning inference compute and inflating identity provider bills before the promotion even ended.

"We were swamped with scripts running datagen prompts, spreading the load across thousands of users. Our only protection was captcha on sign-up, which it seems was not effective."

Robin Fernandes, Engineer at Nous Research

Rebuilding the sign-up and chat flow

Nous used the downtime after the promotion to rework its sign-up and chat flows. Before launching another promotional period, it adopted Vercel BotID Deep Analysis, an invisible captcha that identifies bots impersonating human behavior without disrupting the user experience. BotID was deployed across both the portal and chat interfaces at critical checkpoints:

  • Portal (auth flow): BotID checks run before and after user sign-up or sign-in. Failed checks halt progress immediately.
  • Chat (UI layer): To prevent API-style abuse of the chat app, a heartbeat mechanism triggers periodic BotID checks via tRPC. Successful checks are cached on the backend for a short window, and inference requests only proceed if a recent human verification exists.

That layered setup protects both entry points and continued interaction, balancing security with cost efficiency.

A 3,000% traffic spike gets stopped cold

On October 16 around 12pm AEST, Nous quietly reopened the free tier without marketing promotion. Within days, BotID detected and mitigated a coordinated attack that found the open chat and tried to exploit it.

The traffic spike carried distinct signals:

  • Five to six IPs generated most of the load, with a long tail of smaller sources
  • A majority of traffic originated from Japan
  • 3–4 distinct JA4 hashes correlated across attack attempts

After BotID identified and blocked the traffic, the attack lasted two hours before the attackers gave up. At its peak, Nous saw a 3,000% increase in traffic volume to its chat app. Despite thousands of attempted signups, inference traffic and availability remained stable. Notably, the same JA4 fingerprints BotID marked as "human" at the sign-in portal later appeared as the most-blocked patterns in chat—evidence that attackers adjust their methods, and that deeper behavioral analysis provides adaptive defense beyond surface-level protection.

"We didn't see a spike in inference at all and our chat performance remained steady. The attack burned out once it became clear they couldn't get past BotID."

Robin Fernandes, Engineer at Nous Research

Defense without user friction or ops overhead

By positioning Vercel BotID at multiple points in its user flow, Nous Research stopped automated abuse before it could impact infrastructure, reduced wasted inference spend, and kept the free tier available without compromising security or the user experience. Because BotID responded automatically, the team didn't have to triage traffic issues manually—freeing them to continue hosting an open-source LLM for users worldwide.