AI threats, crawlers, leakage into a model and a fake counterparty
Opselis
Training bots collect store and documentation content. People paste secrets into chat. The scope is crawler policy, DLP and a data boundary in the assistant. Origin stays out of reach, and the question log goes to on-call.
AI traffic is not only ChatGPT in a browser. Crawlers from OpenAI, Google, Anthropic, ByteDance and Perplexity visit documentation, the blog and product cards in millions of requests. Cloudflare AI Crawl Control shows the operator, the path and whether the request got a 200. From there the decision: allow, block, or a stricter robots.txt.
The second vector is internal. An employee pastes a database dump, an API key or a contract into a model. A WAF will not help here. You need an inventory of tools, DLP on mail and endpoints, secrets kept out of chat, and a clear ban on customer data in public models.
The third vector is your own assistant next to the product. Prompt injection and context leakage from a RAG store. The model gets only what the user is allowed to see. The question and answer log goes to the SIEM, with retention for audit.
The fourth is social engineering: an invoice and a “from the board” voice generated by a model. Transfer verification stays in the process, not in intuition. Scope of work: crawler policy at the edge, control of shadow AI, a hard data boundary in the assistant, a playbook for a fake counterparty.