Agent Identity Over API Rate Limits

This is really a unit economics problem disguised as a security problem. In products like Cursor and Perplexity, the expensive part is not one more API call, it is one more successful account that can keep pulling model output through the app. A bad actor can create or rotate many accounts, look like many separate users, and drain free inference across login, browser, and device layers that endpoint rate limits never see.

• Cursor sells plans with included model usage and agent limits, which means abuse shows up as stolen output tokens inside the product, not just excess requests at a public API. The abuse pattern is repeated sign up, use the free pool hard, then cycle identities and devices.
  2 sacra 3 cursor 4 cursor
• The practical fix is identity at the session and device level. Fingerprinting and agent verification let a company connect many accounts back to the same operator, even when IPs, emails, and user agents change. That is why bot and agent identity matters more than simple per key throttling.
  1 sacra 5 stytch 6 stytch
• Perplexity faces the same structural issue as other self serve AI apps. It offers a free product alongside paid tiers and promotional upgrades, so the attack surface is not only scraping but account farming and promo misuse by users trying to get subsidized model answers without paying.
  4 cursor 7 perplexity 8 revolut

The next wave of AI consumer infrastructure will look more like payments fraud tooling than classic API security. As AI apps add browser agents, shopping actions, and higher cost models, the winners will be the ones that can recognize which human or agent is really behind each session and meter access before compute is spent.