AI safety

Lexey has safety checks built in, to protect your business and your customers.

Input filtering

Every customer message is checked before it reaches the support agent. The input filter blocks:

• Prompt injection — Attempts to override the agent's instructions.
• System prompt extraction — Attempts to get the agent to reveal its instructions.
• Abusive content — Offensive or harmful language.
• Adversarial inputs — Messages crafted to manipulate the agent.

A blocked message gets a polite, professional refusal.

Output filtering

Every reply is checked after it has been sent. The checks are recorded for you to review; they don't stop the reply reaching the customer. The output filter flags:

• System prompt leakage — Replies that reveal internal instructions.
• Hallucinated policies — Policies or information that aren't in the knowledge base.
• Inappropriate content — Replies that don't meet content standards.
• Instruction compliance — Replies that suggest a prompt injection worked.

Web content safety

When Lexey reads a web page for you, an AI classifier checks the address and the content before Lexey sees it or anything is saved. Responses from your own APIs, through the API connector, are screened the same way before the agent uses them. If the check can't run, the content is rejected.

Telling customers they're talking to AI

The customer chat always shows that replies come from an AI: an AI marker in the header and a line under the message box. When the agent hands a conversation to your team, the customer sees a notice saying so.