A tool that reads prompts has to be careful with them.
PromptSec sits in front of everything your staff type into an AI tool. That's a lot of trust to ask for, so here is precisely what happens to that data — including the parts still on the roadmap.
Where prompt content goes
Nowhere, by default
Detection runs inside the browser extension. Analysing a prompt takes no network call, so prompt text never reaches our servers unless you explicitly turn on cloud escalation for the ambiguous cases.
Telemetry is metadata
The control plane receives the decision, the detected categories, the tool, and a timestamp — not the prompt. Sensitive spans are masked before anything is transmitted.
Tenant isolation in the database
Per-tenant row-level security is enforced by Postgres itself, not only by application code — a query that forgets its tenant filter returns nothing, rather than someone else's data.
Tamper-evident audit trail
Each audit entry is hashed together with the entry before it. Deleting or editing a row in the middle breaks the chain, and that's detectable on verification.
Cloud escalation, if you turn it on
For prompts the local layers genuinely can't resolve, PromptSec can optionally escalate to a hosted model — Amazon Nova Pro on Amazon Bedrock — for a second opinion. Three things about it are worth stating plainly:
- It's off by default. Nothing escalates unless you switch it on yourself.
- It fails open. If the call errors or times out, the prompt goes through under the local decision — a security feature that breaks someone’s work during a third party’s outage is a feature that gets uninstalled.
- When it's enabled, the ambiguous prompt content is sent to that provider. That's the tradeoff, stated up front so you can decide for yourself.
What we are not claiming
We hold no compliance certifications. Not SOC 2, not ISO 27001, and not certified against the DPDP Act. The architecture is built around data minimisation and on-device processing — the substance those audits look for — but substance isn't a certificate, and we won't imply otherwise.
Detection is not exhaustive. Regex and heuristics catch structured identifiers and known attack shapes well. They still miss things. The named-entity layer that would catch unstructured personal data isn't shipped.
This is one control, not a boundary. It covers browsers. It doesn't cover native apps, terminals, or someone photographing a screen.
Designed for compliance from day one.
Detection methods commoditize. The signed, immutable, regulator-ready log of "who said what to which AI, and when" doesn't. PromptSec is designed to make compliance a byproduct of running the product, not a separate project sitting on top of it. The hash-chained audit trail and detection foundation are live today; public anchoring and signed policy bundles are next.
Pre-built Policy Templates
DPDP Act, RBI cybersecurity, SEBI, IRDAI, and Aadhaar Act templates are designed and ready to fork, customize, and ship.
Signed Policy Bundles
The architecture calls for Cosign/Sigstore-signed policies that the extension checks before applying one. On the roadmap.
Tamper-Evident Audit
Every audit entry is hashed to the one before it, so an edited or deleted row is detectable on verification — live today. Anchoring those hashes publicly to Sigstore Rekor is next.
Regulatory Mapping
Every detection maps to specific Indian regulatory controls — DPDP Act sections, CERT-In directions, RBI/SEBI/IRDAI guidelines. The audit log is structured for how an auditor actually reads one.
Digital Personal Data Protection Act — consent, data fiduciary obligations, breach reporting
Cyber incident reporting within 6 hours, 180-day ICT log retention, time-sync with NIC/NPL NTP
Reserve Bank of India data localization, board-approved cybersecurity policy, incident response
Securities and Exchange Board of India — cybersecurity guidelines for regulated entities, AI in capital markets
Insurance Regulatory and Development Authority — data protection and cybersecurity for insurers
Ministry of Electronics and IT — AI platform labeling, consent for personal data, no bias propagation
Information security management system — globally recognized, adopted by Indian enterprises
Aadhaar number detection is live today, checksum-validated and blocked by default; full mapping to the Act itself is still the design goal
Event export in CSV and JSON is available today, so you can get decisions into your own pipeline before the direct connectors land.
Reporting a vulnerability
Email info@promptsec.in with details and reproduction steps. You'll get a human response, and we won't threaten you for reporting in good faith.
Other ways to reach us