PromptGuard is built with enterprise security requirements in mind. This page outlines our security practices, compliance status, and data handling policies.
At a glance
For security teams evaluating PromptGuard, the three things that usually matter most:- What we do with your data: we scan prompts and responses in real time and never train on them. By default a security event keeps a masked, truncated preview of the content (at most 500 characters); a project’s zero retention setting keeps none. Your LLM provider keys stay with you. (See Data Handling.)
- Proof for auditors: every security decision is recorded in a hash-chained audit log with event ID, threat type, confidence, and timestamp. Verifying it from the dashboard detects a modified event (see Tamper-Evident Audit Trail).
- Where we stand on frameworks: summarized in the table below.
“Compliant” vs “Aligned”: Compliant means we meet the requirement (with a DPA or independent attestation where applicable). Aligned means PromptGuard provides the technical controls a framework requires, but formal third-party certification has not started or is out of scope. We don’t claim certifications we don’t hold.
Security Certifications
“Aligned” means PromptGuard provides the technical controls that satisfy the framework’s requirements. Formal certification (where applicable) requires third-party audit. For enterprise customers requiring specific compliance certifications, contact sales@promptguard.co to discuss your requirements.
EU AI Act Alignment
The EU AI Act (Regulation 2024/1689) imposes requirements on high-risk AI systems, enforceable from August 2, 2026. PromptGuard provides technical controls that map to each requirement:ISO/IEC 42001 Alignment
ISO/IEC 42001:2023 is the international standard for AI Management Systems. PromptGuard’s controls map to its Annex A requirements:Data Handling
What Data We Process
Data Flow
One diagram per surface (proxy, Guard API and SDK auto-instrumentation, the Shadow desktop agent, hosted inference) and the Hybrid and Air-gapped differences are on Data flow.What a security event keeps
PromptGuard scans prompts and responses in memory and forwards them; it does not keep the whole prompt. What it does keep, for each scan, is a security event:- the decision, threat type, confidence and timestamps;
- by default, a masked preview of the content, truncated to 500 characters, plus a content hash. Content from the browser and desktop surfaces is masked before it is stored.
- Zero retention, per project: no preview and no free-text reason is stored.
store_previews: false, per organization.STORE_PREVIEWS=false, for a whole self-hosted deployment (off by default in Air-gapped mode).
Who at PromptGuard can access your data
- Other customers cannot. Every dashboard and API request runs as a database role that row-level security applies to, bound to one Organization per transaction. A query that forgets its filter returns nothing rather than another tenant’s rows. On in production since 17 September 2026.
- Platform administrators can read across tenants. PromptGuard’s back office (the
/internalAPI) runs on a separate, privileged database connection that row-level security does not restrict. Only accounts flagged as platform administrators can use it; the flag is set by another platform administrator or by the service’s configuration. From it an administrator can list and manage customer accounts and plans and see system-wide blocked events: user, threat type, confidence, reason, model and source IP. That view does not include the masked preview. Plan changes made there are written to the audit log. - Infrastructure access is held by the founders. No one else at PromptGuard has access to production. Anyone with database access can read stored events, including masked previews. Infrastructure accounts (Google Cloud, Supabase, GitHub) require MFA. There is no just-in-time elevation or break-glass tooling: access is standing, not requested per incident.
- We do not train models on your data.
Delete your data
What deletion does not remove:
- The audit log. Audit events are never purged. When their Organization is deleted they are kept, unlinked from it, so the record of what happened (including the deletion itself) survives.
- Events in Organizations you don’t own. Deleting your account removes your link to them; the Organization’s owner keeps the events.
- Backups. The database provider’s automated backups keep deleted rows until those backups expire.
- Billing records held by Stripe.
Infrastructure Security
Cloud Infrastructure
Encryption
Network Security
- All endpoints require HTTPS
- Cloud Armor anti-abuse rate limiting at the network edge
- No public SSH access to infrastructure (Cloud Run has no hosts to log in to)
- The API reaches the database over TLS through Supabase’s connection pooler
Access Control
Authentication Methods
Signing out ends the session at once on every server; a password change ends every other session, and a password reset ends them all.
API Key Security
- Keys are stored only as a one-way SHA-256 digest
- Only the prefix (
pg_live_xxxxxxxx...) is stored in plain text - Full key shown only once at creation
- Keys can be rotated without downtime
- Project-scoped API keys (no per-key permission types; tier gating via subscription)
Role-Based Access (Enterprise)
Audit Logging
What’s Logged
Accessing Audit Logs
- Go to Dashboard → Organization → Audit Log
- Filter by date range, event type, or user
- Export as CSV or JSON
Log Export (Enterprise)
Enterprise customers can configure:- SIEM Integration: signed webhook deliveries and pull APIs for Splunk, Microsoft Sentinel, Datadog and others (SIEM export)
Incident Response
Security Incident Process
- Detection: Automated monitoring + manual review
- Containment: Isolate affected systems
- Investigation: Root cause analysis
- Notification: Affected customers notified within 72 hours
- Remediation: Fix deployed, post-mortem published
Reporting Security Issues
Found a vulnerability? Email security@promptguard.co. Response times, scope, safe harbour and how we credit researchers are in our vulnerability disclosure policy.Where Your Data Is Processed
Cloud
Both are in Northern Virginia, US. Data flow lists what
each one stores. There is no hosted EU or APAC region, and we do not
offer a choice of region.
Self-hosted
To keep data in a particular country or in your own infrastructure, self-host: run the engine on your servers in Hybrid mode, or the whole stack in Air-gapped mode. Where your data is processed is then a property of where you deployed it. See Self-Host vs Cloud for what each mode sends to us, Hybrid forwards no prompt content by default, and masked previews only if you setFORWARD_MODE=content.
Vendor Security
Subprocessors
The list of subprocessors, with what each receives, where it runs and whether it is optional, is kept in one place: promptguard.co/subprocessors. On Cloud it includes Hugging Face, which runs the machine-learning and LLM detection tiers and receives the text being scanned when a scan reaches those tiers. A self-hosted engine with local inference calls no one.LLM Providers
PromptGuard forwards requests to the LLM provider you configure, such as:- OpenAI
- Anthropic
- Google AI
- Cohere
- AWS Bedrock
- Azure OpenAI
AI Agent Governance
PromptGuard provides three governance capabilities for AI agents operating in production:Behavioral Drift Detection
After an agent accumulates sufficient observations, PromptGuard freezes a behavioral baseline capturing the agent’s normal tool-usage distribution. Every subsequent request is compared against this baseline using Jensen-Shannon divergence. If the distribution shifts beyond the configured threshold, aBEHAVIORAL_DRIFT alert fires.
Tamper-Evident Audit Trail
Every audit event’s SHA-256 hash covers every field the event stores and incorporates the previous event’s hash in your organization’s chain, which every PromptGuard instance extends in sequence.POST /dashboard/audit-log/verify-chain verifies your organization’s events (an admin) or your own (anyone else) over any time range. It recomputes each event’s own hash, so it detects an event whose fields were modified; the response says so in detects: "modification". It does not walk the links between events, so it cannot show that an event was deleted. The evidence package’s Audit Chain Integrity control (CC7.4) reports the same checks.
If you exported an evidence package before 3 September 2026, re-export it. Two things changed:
- Audit events written before 27 August 2026 were signed over eight of their fields (hash version 1). Events since then are signed over every stored field (hash version 2). Version-1 events are never re-signed and are reported as unverifiable, not intact, not tampered. Every package now states
v1_count,v2_countandcutover_atso you can see which rows carry which signature. - Earlier packages could show CC7.4 as compliant over a period in which no events were verified. A period with nothing to verify now reports
insufficient-datawith the narrative “No events in window; chain not evaluated”, and a package is not produced at all if the audit log cannot be read. - A row whose category, severity or outcome predates a schema narrowing can no longer be decoded, and is reported rather than treated as a read failure: it is excluded from
verified_countand counted inmalformed_count, with the affected event IDs listed (up to 20) inmalformed_events. It does not make the chain around it look broken.
Governance Reports
Generate auditor-facing narrative reports covering these governance capabilities:/dashboard/* takes a dashboard sign-in, not an API key. These endpoints are the API the dashboard at app.promptguard.co calls. They are not part of the public API, and no API key or long-lived token works on them. Everything under /api/v1 takes your X-API-Key instead (API reference).To script one anyway, sign in with your email and password. This needs an account that has a password, and an email domain that requires SSO is refused. The access token lasts one hour.YOUR_SESSION_TOKEN or YOUR_SESSION_COOKIE, it means the access_token from step 1, sent as Authorization: Bearer <token> or as the access_token cookie. A request acts on your personal organization unless it sends X-Organization-Id.Scheduled Compliance Reports
PromptGuard can email a compliance summary to the people you list, every week or every month. Set one up on Dashboard → Compliance Report, in the Scheduled Reports card, with Schedule Regular Reports:- Frequency: Weekly or Monthly
- Frameworks: one or more of SOC 2, GDPR, HIPAA and ISO 27001
- Recipients: comma-separated email addresses, at least one. Your own address is filled in to start.
What the email contains
The subject is[PromptGuard] Weekly Compliance Report - <date> (or
Monthly). The email gives:
- Audit events, Threats flagged and Critical events for the period. The plain-text version also gives the block rate.
- Control coverage for each framework you chose: how many of its controls the period produced any evidence for, out of how many, as a percentage. It is worked out from the same control table as the Compliance evidence page. Like that page, it says how much evidence exists, not whether a control passed, and it is not an attestation.
- A link to the Compliance page.
When it is sent
- The first report goes out on the next hourly delivery run after you create the schedule.
- After that, a weekly schedule is sent again once 7 days have passed since its last send, and a monthly one once 30 days have passed.
- Every report covers the last 30 days, whichever frequency you choose, so consecutive weekly reports overlap.
- A paused schedule sends nothing until you resume it.
- If the email reached none of the recipients, the schedule is not marked as sent, and the next hourly run tries again.
Enterprise Security Features
Available on the Enterprise tier:Security Questionnaire
Need to complete a vendor security assessment? We answer SIG Lite, CAIQ Lite and your own questionnaire on request, and can send an assurance pack covering tenant isolation, encryption and release evidence. There is no SOC 2 report or penetration-test report to send. Contact security@promptguard.co.Next Steps
Security Overview
Learn about threat detection
Audit Logs
Monitor user activity
Enterprise
See Enterprise features
Contact Sales
Discuss your requirements