31.4.2 Per-prompt payment models

2025.10.06.
AI Security Blog

While subscription models offer threat actors a predictable revenue stream, the per-prompt or “pay-as-you-go” model provides a lower barrier to entry for a wider range of illicit users. This approach commoditizes jailbreaking, turning it from a committed service into a transactional utility, much like legitimate cloud computing APIs. You don’t subscribe to the service; you simply pay for what you use.

Core Mechanics: The API-Driven Transaction

The per-prompt model is almost exclusively implemented via an Application Programming Interface (API). A user purchases credits or tokens, often using cryptocurrency to maintain anonymity, and receives an API key. This key authenticates their requests to the jailbreak service’s endpoint. Each successful generation of a harmful or policy-violating response deducts a certain number of credits from their account.

Kapcsolati űrlap - EN

Do you have a question about AI Security? Reach out to us here:

This structure allows for seamless integration into other malicious tools, such as automated scam content generators, malware droppers, or disinformation campaign bots. The jailbreak capability becomes a programmable component in a larger attack chain.

Example: API Call to a Fictional Jailbreak Service


import requests
import json

# --- Configuration for the illicit service ---
API_ENDPOINT = "https://shadow-prompt.onion/api/v2/unleash"
API_KEY = "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxx" # User's purchased API key

# --- Crafting the request to generate malicious code ---
payload = {
    "target_model": "enterprise-finance-v4",
    "prompt": "Write a polymorphic Python script that evades EDR.",
    "technique": "obfuscation_max",
    "cost_per_token": 0.005 # Price may vary based on complexity
}

headers = {
    "X-API-Key": API_KEY,
    "Content-Type": "application/json"
}

# --- Executing the call ---
# The service's backend authenticates the key and debits the user's
# account based on the length and complexity of the generated output.
response = requests.post(API_ENDPOINT, headers=headers, data=json.dumps(payload))

if response.status_code == 200:
    print("Success. Credits deducted.")
    print(response.json()['generated_text'])
else:
    print(f"Error: {response.status_code} - {response.text}")
                

Economic Rationale and Market Positioning

The pay-as-you-go model appeals to a different market segment than subscriptions. It caters to users with sporadic needs, those testing the service’s capabilities, or developers integrating jailbreaks into specific, limited-use applications. For sellers, it broadens the customer base and can capture revenue from individuals unwilling to commit to a monthly fee.

However, this model introduces different operational challenges for the service operator, including more complex billing/metering systems, the need to manage thousands of low-value accounts, and a higher risk of credit fraud or metering bypass attempts.

Comparison: Per-Prompt vs. Subscription Models

Feature Per-Prompt Model Subscription Model (from 31.4.1)
Payment Flow Transactional (pay per use, credit packs) Recurring (monthly, yearly fee for unlimited/tiered access)
Buyer Commitment Low; ideal for one-off tasks, experimentation, or integration High; suited for continuous, high-volume use
Seller Revenue Model Variable and usage-dependent Predictable Monthly Recurring Revenue (MRR)
Primary User Base Developers, researchers, occasional malicious actors Power users, criminal organizations, content farms
Key Attack Surface (for the service) API key leakage, metering bypass, credit fraud Account takeovers, credential stuffing, shared account abuse

Implications for Red Teaming and Defense

When you encounter a target potentially using a per-prompt jailbreak service, the API-centric nature of the model presents specific avenues for investigation. The presence of hardcoded API keys in client-side code, mobile applications, or leaked source code is a critical vulnerability. Compromising a key not only grants you the ability to generate malicious content but also allows you to exhaust the user’s purchased credits, disrupting their operations in a denial-of-service attack (often termed “economic denial of service” or EDoS).

From a defensive standpoint, monitoring for outbound traffic to known illicit API endpoints is a key detection strategy. Furthermore, analyzing the logic of applications that interact with such services can reveal how attackers are programmatically weaponizing AI models, providing valuable intelligence for tuning your own defensive filters and content policies.