Your inference provider or router says prompts may have been exposed: find what was in them, rotate secrets, and tighten routing
For teams that send LLM traffic through routers or third-party inference hosts and get a notice like Novita's, Phala's or OpenRouter's. The prompt is the breach: whatever you or your agents put in it is what leaked.
Scope
Match the notice to your traffic.
- Pull router/provider logs for the model, provider and time window named in the notice
- List which apps, agents and API keys sent that traffic
- If you log prompts, export the affected requests; if you don't, list what each app normally puts in its context
Find secrets
Agents leak more than people do.
- Scan exported prompts and tool outputs for API keys, tokens, passwords, connection strings and private repo URLs
- Include secrets agents read from files, environment or shell output and passed back to the model
- List personal and customer data in the prompts for privacy review
Rotate and watch
Assume anything found was copied.
- Rotate every secret found, starting with cloud and source-control credentials
- Check those accounts for use since the window opened (cloud API calls, new IAM users, repo clones)
- Rotate the router/provider API key itself
Tighten routing
Control who processes your prompts.
- Allowlist vetted providers and disable automatic fallback for sensitive apps
- Don't rely on a zero-data-retention label; get it in the contract
- Plant canary tokens in test traffic to detect exposure
Close out
Make it the default.
- Move agents to short-lived scoped credentials from a broker
- Add secret scrubbing before prompts and tool outputs leave your network
- Record the notice, scope and rotations for compliance and customer questions