Check every request to the chatbot's API route for automation before your server calls the model, so a bot's message is refused before it costs any tokens. Rate limits per IP address and per session help against simple scripts, but a bot spread across residential proxies and fresh sessions stays under them. Spending caps stop the bill, but they stop the chatbot for real customers too. Prosopo Protect runs as a script tag on the pages with the chat widget, with edge protection on Cloudflare Workers or AWS Lambda@Edge in front of the chat API, and names the reason for every block.
AI chatbot bot protection: stop bots burning your LLM credits
Competitors, resellers and scripts can run up your model bill through the chatbot on your site, and spending caps only stop the bill by switching the chatbot off for real customers as well. Prosopo Protect checks every request to your chat API before it reaches the model, so a bot's message is refused before it costs you any tokens.
Who is burning your chatbot's credits
Every message to an AI chatbot is a paid call to a language model. A real customer who asks three questions costs you a few cents. Anyone who can script your chatbot can make it cost you far more, and they pay nothing for it.
Competitors running up your bill. A script that holds long conversations with your chatbot, all day, turns your model budget into their weapon. OWASP calls this denial of wallet and lists it under LLM10:2025 Unbounded Consumption in its Top 10 for LLM applications.
People using your chatbot as a free general-purpose AI. In December 2023 users found that the ChatGPT-based chatbot on a Chevrolet dealer's website would write Python code and answer questions unrelated to cars. GM Authority reported that the dealer turned it off. It still works on many sites: in December 2025 a Hacker News commenter found that Amazon's Rufus shopping assistant and the chat widget on their city government's site would both write code on request, and wondered whether anyone had built a service to spread a coding assistant's traffic across chat widgets like these. One person doing this costs little, but a script doing it all day costs you a model call for every answer.
Resellers calling your chat API directly. Your chat widget calls an API route on your server, and that route calls the model with your key. A reseller skips the widget, calls the route from their own servers and sells the answers on. This is the same pattern as the competitor in our scraped SaaS case study, who consumed a free tier through its API and resold the results.
There is open-source tooling for it. gpt4free reverse-engineers the AI features of other companies' sites and exposes them as one free API. It had 67,080 GitHub stars on 18 November 2026 and was still being updated. OpenAI sent it a takedown request in April 2023, and sites it used, such as Phind, had to ask to be removed. When the takedown was discussed on Hacker News, one commenter described it as using "someone else's OpenAI credentials so that you don't have to pay for it". Another, in a later thread, said that if you run any freely accessible LLM, someone will try to exploit it for profit, and "it's usually pretty obvious when they find it because your bill explodes."
The market for resold model access is large enough to attract organised groups. In November 2026 Google's threat intelligence group reported a resale service, Jet Night, that made more than 500,000 model calls in under eight hours on one victim's stolen keys and passed nearly $1M in token costs to them, while selling "unlimited" access for about $60 a month. Stolen keys are a credential problem, but your chat endpoint does the same job as a key, and you hand it to every visitor.
Scripts extracting what your chatbot knows. A chatbot grounded in your documentation, catalogue or pricing will answer questions about all of it. A script asking thousands of questions copies that knowledge out, one paid answer at a time. OWASP lists this under the same risk as model extraction.
Fake accounts farming free credits. If your product gives each new account free AI credits, every automated signup is a grant of credits to a bot. That is a fake account creation problem with a model bill attached.
Why bot conversations cost more than they look
A chat API sends the model the whole conversation every time, so each turn costs more than the last. A bot that keeps a conversation going is the most expensive visitor you have.
Take a chatbot with a 3,000-token system prompt and retrieved context, a 50-token question and a 400-token answer on each turn, at illustrative prices of $3 per million input tokens and $15 per million output tokens.
| Conversations | Input tokens | Output tokens | Cost |
|---|---|---|---|
| 1,000 single questions | 3.05 million | 400,000 | $15 |
| 1,000 conversations of 20 turns | 146.5 million | 8 million | $560 |
| 1,000 conversations of 20 turns a day, for 30 days | 4.4 billion | 240 million | $16,800 |
A 20-turn conversation costs 37 times as much as a single question, because by the last turn the model is reading almost 12,000 tokens to answer 50. A bot can push the cost higher by asking for long, thorough answers, since output tokens cost five times as much as input tokens in this example, and it can drive the conversation with a cheap open model of its own. Prompt caching lowers the price of the repeated part, which helps, but you still pay for every turn and the bot still pays nothing.
One cost comes from your own design. A Hacker News commenter described a nonprofit that got a $2,000 model bill in a month with very few conversations, because the chatbot called the model to greet every visitor on every page load. If your widget does that, every crawler and bot that loads a page costs you a model call. Wait for a person to send a message before calling the model.
What the usual defences stop, and what gets through
| Defence | What it stops | What gets through |
|---|---|---|
| Rate limits per IP address | Fast scripts from a few addresses | Bots spread across residential proxies at a human pace |
| Per-session or per-conversation caps | One very long conversation | Bots that start a fresh session for each conversation |
| Daily spending caps | A runaway bill | Nothing, because once abuse reaches the cap the chatbot stops for real customers too |
| Requiring sign-in | Anonymous scripts | Bots using fake accounts, which are cheap to make |
| A CAPTCHA before the chat opens | Bots that use the page | Bots calling the chat API directly, and it puts a puzzle in front of every customer |
| Bot detection on every chat request | Automation on the page and on the API, judged on behaviour, connection and whether the browser's claims hold together | Very little, if it covers the API route and runs before the model is called |
Keep the first four. They are cheap and they remove the bots that do not try. But the people reselling model access already work around them: Google's report describes one group that repeatedly created and deleted short-lived accounts to stay under rate limits, while watching the victim's bill to track the damage. A spending cap in particular is a sensible backstop. But each of them either counts something a bot can change or limits everyone at once. None of them tells you whether the visitor sending the message is a customer.
How Prosopo Protect stops chatbot abuse
Prosopo Protect is a script tag on the pages that carry your chat widget, with edge protection on Cloudflare Workers or AWS Lambda@Edge in front of the API route the widget calls. It checks every chat request before it reaches your server, so a blocked message never triggers a model call and costs you no tokens.
- The chat API covered, not just the page. Blocked API calls get a 403 with an
X-Prosopo-Decisionheader, or a 401 that protect.js handles in the browser when a challenge is needed. A reseller calling the route from a server is refused, and your widget can show a challenge to a suspect session in the page. - Detection that doesn't depend on request rate. A bot holding one conversation every few minutes from a home broadband address looks normal on volume. Protect looks at how the visitor behaves, whether the browser contradicts itself and whether the connection shows signs of a relay, which slowing down cannot fix.
- Residential proxies caught on the connection. Residential proxy detection checks the TCP connection for signs of a relay and matches exit addresses against proxy networks we rent and probe ourselves, so spreading a bot across home addresses does not hide it.
- Customers keep chatting. Most real visitors never see a challenge, and suspicious sessions can be challenged with an image, puzzle or proof-of-work rather than blocked. In the scraped SaaS case study, 79% of real visitors were cleared without seeing a puzzle.
- Patterns become standing rules. Prosopo manages the detection rules. Once a bot's pattern is established it is turned into Access Rules on the networks it comes from, and you can add your own on IP range, ASN, country, user agent, JA4 TLS fingerprint and IP category.
- AI agents on your terms. Choose whether verified AI agents are allowed, challenged or blocked across your account, so an agent acting for a customer is not caught by accident.
- Fast enough for chat. On a production ticketing deployment the verification hop measured 39.3 ms at p95 over 30 days, which is small next to the time a model takes to start answering.
Every block names the reason it was made, so when a customer says the chatbot refused them you can see why.
What it costs
Protect is priced by number of sites rather than per request, within fair use, from $1,990 a month. Your model bill grows with every message a bot sends; what you pay Prosopo depends on your sites instead. Every deployment starts with 14 days in preview mode, so you see which chat sessions it would block, and what they were costing you, before it blocks anything.
Prosopo products for protecting AI chatbots
- Enterprise (Prosopo Protect) — check every chat request on the page and the API before the model is called.
- Residential Proxy Detection — catch bots renting home broadband addresses to stay under rate limits.
- Access Control — block the hosting networks, TLS fingerprints and user agents a bot uses.
- Bot Protection — a 0–1 score per request so your backend can shorten or refuse a suspect conversation.
For the wider picture of AI businesses, see bot protection for SaaS and AI companies.
See what's reaching your site
A Bot Exposure Assessment is a 30-minute call with an engineer. We look at your traffic, show where your current setup leaves gaps, and agree how a trial would be judged. Nothing to install first.
Questions
Something else? Ask an engineer.
How do I protect my AI chatbot from competitors' bots burning my credits?
What is denial of wallet?
Denial of wallet is an attack that aims to run up a service's bill rather than take it offline. OWASP lists it under LLM10:2025 Unbounded Consumption in its Top 10 for LLM applications: attackers send a high volume of requests to exploit the cost-per-use pricing of AI services. Against a chatbot it means scripted conversations that each cost you model tokens and cost the attacker nothing.
Why don't rate limits stop chatbot abuse?
A rate limit counts requests from one IP address, session or account. A bot that sends each conversation through a different residential proxy address, starts a new session for each one and keeps a human pace stays under every limit you can set without limiting real customers. Rate limits are still worth having, because they stop simple scripts cheaply, but they cannot tell a slow bot from a customer.
Should I put a spending cap on my chatbot?
Yes, as a backstop. A daily cap on model spend stops a runaway bill. But when abuse reaches the cap, the chatbot stops answering everyone, including the customers it is there for, so an attacker who wants your chatbot offline only has to reach it. A cap limits the damage; it does not tell bots and customers apart. Blocking the bots before the model is called means the cap is rarely reached.
Will a CAPTCHA stop bots using my chatbot?
A CAPTCHA in front of the chat window stops bots that use the page, but it adds a puzzle before every conversation, which is the opposite of what a support or sales chatbot is for. Bots that call your chat API directly never see it. Prosopo Protect checks the visitor in the background and only shows a challenge to sessions that look automated, and it checks the API calls as well as the page.
Can someone use my chatbot as a free ChatGPT?
Yes, and it has happened in public. In December 2023 people found that the ChatGPT-based chatbot on a Chevrolet dealer's website would write Python code and answer questions that had nothing to do with cars, and the dealer switched it off. A resale operation goes further: it calls your chat API from its own servers and sells the answers on, so you pay for the model and it keeps the revenue.
Is chatbot abuse the same as LLMjacking?
No, but they feed the same market. LLMjacking means stealing API keys or cloud credentials and using them to call a model directly; Google's threat intelligence group reported in November 2026 that one resale service passed nearly $1M in token costs to a single victim. Chatbot abuse needs no stolen credentials: the bot uses your chat endpoint, which calls the model with your key on behalf of anyone who sends a message. Key rotation and spending alerts deal with LLMjacking. Checking each chat request for automation deals with chatbot abuse.
Does blocking bots on my chatbot block AI agents my customers use?
Only if you choose that. Prosopo Protect lets you allow, challenge or block verified AI agents as a setting across your account, so an agent acting for a customer can be let through while unverified automation is stopped.
How much does chatbot bot protection cost?
Prosopo Protect is priced by number of sites rather than per request, within fair use, from $1,990 a month, so what you pay Prosopo depends on your sites rather than on how many messages a bot sends. It starts with 14 days in preview mode, so you see what it would block before it blocks anything.
Other attacks Prosopo stops
- Account takeover prevention: stop takeover bots at login Account takeover prevention: the five routes into an account, the warning signs, and how to stop the automated ones at login, password reset and account recovery before a password is checked.
- Stop Black Friday Sale Automation with Prosopo Stop hoarding bots, card testers and checkout scripts during Black Friday sales — drops into cart, checkout API or CDN edge without slowing real shoppers.
- Stop Click-Through Rate Fraud with Prosopo Stop click-farms and bot clicks polluting ad spend. Prosopo verifies clicks at your tracking endpoint, ad server or CDN edge — metrics reflect real users.
- Credential stuffing prevention: stop breached-password replay at login Credential stuffing prevention and protection for your login form, auth API and password reset. Stop breached-password replay at the edge; real users pass.
- Stop Denial of Inventory Attacks with Prosopo Stop hoarding bots locking up carts and reservations. Prosopo blocks add-to-cart abuse at checkout API or CDN edge — real buyers get the stock.
- Stop Loyalty Programme Automation with Prosopo Stop fake signups and points-farming bots at signup, claim and redemption endpoints — real members earn and redeem unimpeded, backend or CDN edge.






