← All capabilities New
PromptGuard Security
Agentic analytics only works if untrusted prompts cannot steer the model into leaking data or bypassing policy. PromptGuard sits in front of chat, completion, and agent calls.
Requests are inspected for injection, jailbreak, role-play, and exfiltration patterns before they reach the LLM. Heuristics catch paraphrases that a static denylist would miss.
Blocked calls fail closed with an audit-friendly error so security and data teams can review what was attempted without executing it.
What this does for the business
- Let employees ask questions of sensitive datasets without opening a prompt-injection hole.
- Keep CodeAgent and LangGraph runs inside the same security envelope as chat.
- Show customers and auditors that AI traffic is screened, not trusted by default.
API
POST /api/ai/chat POST /api/ai/completion Full reference: api.dsnresearch.com/docs