Secret Leak Detector watches the surfaces where secrets accidentally land: log lines, prompt context, screenshot uploads, support attachments, decision bundles. When a candidate match shows up, the detector routes the finding to the right team, opens an automatic rotation request for the impacted credential, and bundles the evidence for the post-incident review. The goal is rotation in minutes, not days.
The detector runs against five surfaces. Log lines (every entry shipped to your log backend gets scanned in flight). Prompt context (every input the AI fleet reads passes through the detector before reaching the model). Decision bundles (every saved bundle is rescanned). Support attachments (screenshots, files, error reports uploaded by humans). Configuration blobs (saved configs and runbooks). One detector, one alert path, no fragmented per-surface tooling.
The detector groups findings into broad categories: cloud-provider credentials (AWS, GCP, Azure), platform tokens (GitHub, GitLab, Slack, Stripe, OAuth), private keys, JWTs, and high-entropy values that look credential-shaped. The detection rules themselves are intentionally not published. A public rulebook is an attacker's evasion guide; a private rulebook keeps the defense effective. Customers see categories and counts, not regexes.
A finding does more than fire an alert. The detector identifies the credential's owning service (when known) and opens a rotation request through your secrets backend (Vault, AWS Secrets Manager, GCP Secret Manager, Azure Key Vault). The owning team is paged with the rotation request pre-filled. Approve and the rotation runs. The whole loop is designed to compress the leak-to-rotation time, which is the number that actually matters.
Every finding writes to the same hash-chained ledger as Agent Ledger and Decision Bundles. A weekly report shows finding volume per surface, time-to-rotate distribution, false-positive rate, and trend. The data is part of the standard SOC2 / HIPAA / ISO 27001 export, so audit reviewers see the full picture without a separate request. False positives marked by an operator feed back into a private tuning loop; the categories shown to customers do not change but the underlying precision does.
Subscribe to Nova AI Ops on YouTube for demos, tutorials, and feature deep-dives.
A leaked credential rotated within an hour is a near-miss. The same credential rotated next quarter is a breach. The detector exists to keep findings on the right side of that line.