Documents, web pages, emails and tool outputs can carry text written to make your AI do something you did not ask for. AI Cordon checks that text before your system acts on it and shows where the injected part is.
One API key, three models. Real-time requests or batch files of up to 50,000 documents.
Prompt injection detection for everything your AI system reads: retrieved documents, browsed pages, emails, tool and agent outputs.
Answers your yes/no questions about a document with a probability, so routing, filtering and policy checks become a single API call.
Multilingual text embeddings for semantic search, clustering and retrieval, through the OpenAI-compatible embeddings endpoint.
Pay per token. Batch requests cost half the real-time price.
| Model | Real-time, USD per 1M tokens | Batch, USD per 1M tokens |
|---|---|---|
| Intent | 0.03 | 0.015 |
| Intent spans | 0.09 | 0.045 |
| Decider | 0.015 | 0.0075 |
| BGE-M3 | 0.01 | 0.005 |
Intent spans are billed only for documents flagged as injections, on the tokens read to locate the injected text. Decider bills the document and your questions; its answers are free.
Create a key in the console and send a document.
curl https://api.thetabench.com/v1/detect \
-H "Authorization: Bearer $AI_CORDON_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "intent", "text": "…", "spans": true}'