Security

Plain-English version of how Agent-Shield handles your data, what we can and can't see, and what's open for you to inspect.

What happens to content you scan

• Your content is processed in memory, on Cloudflare's edge, and the raw text is never written to disk, never logged, never retained.
• What leaves the scan: a verdict, per-class scores, and findings shaped as { class, match_count } — never matched text. We tell you that a secret pattern fired and how many times, never the secret itself.
• Each response carries a content_hash: a keyed HMAC of your content, used only so you can correlate a scan with your own logs. It can't be reversed into your content, can't be brute-forced without our key, and the same content scanned by two customers produces different hashes.
• Optional signed receipts attest that a scan happened (tamper-evident proof-of-scan). They are not audit-grade: no append-only log yet, HMAC not Ed25519 yet. We say so on the receipt itself.

What we store

• API keys and tier (KV): the key string, tier, label, creation date. Email only if you give it — signup works without one.
• Usage counters (KV): per-key daily scan counts for quota enforcement, expiring after 48 hours.
• Billing (Stripe): handled entirely by Stripe Checkout; we never see or store card numbers.
• Nothing else. No scan content, no prompts, no tool outputs.

What's open, what's closed

• Open (GitHub): the Cloudflare Worker — the full scanning engine, detectors, and API. Run it yourself; the code is the documentation.
• Closed: the private hold-out evaluation set, billing internals, HMAC keys. The future classifier weights stay private until calibration is done — then we'll say so explicitly.

Subprocessors

• Cloudflare — edge compute (Workers) and key-value storage. Scans run in the Cloudflare region nearest you.
• Stripe — subscription billing.

Responsible disclosure

Found a vulnerability? Email StarTek@Startekenterprises.com. We aim to acknowledge within 48 hours. Please don't probe other customers' keys; free-tier abuse gets rate-limited.

What this page doesn't claim

We publish our detection benchmarks — including the misses — because a security product that won't show you its false-positive rate hasn't earned your trust. This page describes handling, not guarantees: no scanner catches everything, and ours tells you exactly where it stands. See /benchmarks.