Article
How Bella protects your AI workspace before model calls
Last updated: June 22, 2026 · 3 min read
Bella is chat-first. That means a user can ask for CRM work, commission checks, case notes, and follow-ups from one message. Because chat is powerful, Bella now checks risky input before it reaches any AI provider.
What happens before the model
- Bella receives the message.
- A local deterministic gate checks for unsafe patterns.
- If the message is unsafe, Bella returns a safe refusal.
- If the message is normal insurance work, Bella continues to memory, tools, and model reasoning.
What Bella blocks
- Prompt-injection attempts that ask Bella to ignore hidden instructions or safety rules.
- Requests to reveal API keys, secrets, passwords, tokens, private keys, or environment files.
- Malware, phishing, credential theft, fake-document, fraud-evasion, weapon-harm, and self-harm instructions.
- Oversized payloads that should not be forwarded to a model provider.
What Bella stores
Bella does not store the raw blocked prompt. The audit record stores a hash, a redacted preview, the safety code, and tenant/user metadata. Secrets, bearer tokens, SSNs, IDs, card numbers, dates of birth, emails, and phone numbers are scrubbed from the preview.
What still works
Normal insurance language is allowed. Bella should not block everyday requests like generate an insurance card, collect a driver license copy, flag possible fraud, prepare audit notes, explain password reset, review policy rules, or clean up duplicate leads.