Risk classifier sensitivity

The risk classifier inspects outbound prompts for things that shouldn't leave — secrets, personal data, prompt-injection attempts, and exfiltration-shaped tool sequences.

Policy → AI risk classifier sensitivity controls how aggressively it acts:

• Low — suppresses low-confidence signals, for a quieter feed.
• Medium (recommended) — records medium- and high-severity findings as warnings.
• High — the strictest setting.

Findings appear on the Activity screen either way; sensitivity changes what is acted on, not what is recorded.

Secrets are blocked regardless

A detected credential — an AWS key, a private key, a provider token — is blocked in the request path independently of this setting. Sensitivity governs the softer signals.

Reading the result

On Activity, a request with findings shows chips in the Signals column (aws_access_key_id, email_address ×3, and so on) and its outcome is Blocked if policy stopped it. Click through to the conversation when prompt capture is on.