Security, Privacy & Governance
DoCoreAI is built for production GenAI systems operating under security, risk, and governance constraints. We observe system behaviour — never model content. By design, we never store prompt or output content; only the telemetry needed to compute cost, efficiency, and ROI.
- Content-blind by design. We never see or store prompts, responses, or customer payloads.
- Non-blocking architecture. DoCoreAI never sits in your LLM request path — your application is unaffected if governance is unavailable.
- PII hard-blocked at the edge. Prompts containing email addresses, SSNs, credit card numbers, phone numbers, or IP addresses are blocked before the LLM call is made.
- Tenant-isolated. Your telemetry is never visible to other customers, accessible only via your own token.
- Metadata stored in Manchester, UK. UK GDPR compliant. Local telemetry never leaves your environment.
- SOC 2 & ISO 27001 alignment are on our compliance roadmap — SOC 2 audit targeted Q3 2027.
- Configurable retention, with telemetry opt-out and data deletion supported on request.
01Trust Philosophy
DoCoreAI follows a minimal, privacy-first design. We collect only what's needed to understand how your GenAI systems behave in production — without ever accessing or storing prompts, responses, or customer content.
This lets your team operate, monitor, and govern GenAI systems without introducing new privacy or security risk of its own.
02What We Collect — and What We Don't
DoCoreAI is intentionally content-blind. Here's exactly where the line is drawn.
- Token counts (prompt and completion)
- Request latency, retries, and error signals
- Model and version identifiers
- High-level usage and cost metrics
- Prompt text or prompt bodies
- Model responses or generated content
- User inputs or application payloads
- LLM provider API keys or secrets
- Personally identifiable information (PII)
03Data Flow Overview
Application content flows directly between your system and your LLM provider — completely unchanged. DoCoreAI captures only metadata through a separate, telemetry-only channel inside your application process.
(direct, unchanged)
No prompts. No outputs.
DoCoreAI does not proxy, intercept, or store LLM requests or responses. The SDK intercepts the call only to extract metadata — model name, token count, latency, and cost — before passing the call through to the LLM provider unchanged. No prompt content or response content is read, processed, or transmitted to DoCoreAI servers.
04PII Detection
DoCoreAI includes an optional PII detection layer that scans outgoing prompts before the LLM call is made — catching sensitive data at the source, not after it has been logged somewhere downstream.
- Email addresses
- Social Security Numbers (SSNs)
- Credit card numbers
- Phone numbers
- IP addresses
- Hard block — the LLM call is never made
- No redaction, no pass-through with warning
- Alert written to the developer console
- No PII value is stored anywhere
- Block happens entirely in-process
IP addresses are treated with the same detection confidence as SSNs and credit card numbers. Technical prompts containing IP addresses — such as log analysis queries or network debugging tasks — may trigger the PII filter and be blocked. This is intentional conservative behaviour. Known-safe patterns can be configured via the governance config to prevent false positives in technical workflows.
05Data Residency
Understanding exactly where your metadata goes — and where it stays.
A local SQLite database is created in your environment at install time. All raw telemetry — per-request token counts, latency, cost, and governance decisions — is written here first. This data never leaves your environment. It is under your full control and can be deleted at any time.
Aggregated metadata is synced to the DoCoreAI cloud dashboard, hosted on infrastructure located in Manchester, England (United Kingdom) via Hostinger. The UK operates under UK GDPR — substantially equivalent to EU GDPR. No prompt content, response content, or PII is ever included in this sync.
Teams with specific data residency requirements outside the United
Kingdom — including EU-only, US-only, or region-specific constraints
— should contact us before deployment to discuss available options.
Enterprise plans support custom deployment configurations.
Contact us about data residency →
06Security Controls
All communication between the SDK and the DoCoreAI cloud uses HTTPS/TLS. We store only non-sensitive operational telemetry — numeric metrics, timestamps, and model identifiers. No customer content or secrets are ever stored.
DoCoreAI monkey-patches your LLM SDK at the Python process level. It intercepts calls after your SDK has handled authentication — it never reads, stores, or transmits your LLM provider API keys. Your keys stay inside your SDK client object, in your runtime, under your control.
Each customer's telemetry is logically isolated and accessible only through their own token and authenticated account. Customers cannot view telemetry belonging to other customers.
07Risk & Reliability Design
DoCoreAI never sits in the execution path of your LLM requests. If the DoCoreAI cloud dashboard is unavailable, local governance continues to operate independently. Your LLM calls are never blocked or delayed by DoCoreAI's availability status.
Telemetry collection is designed to fail safely — without impacting model execution or your application's availability. A failure in the telemetry sync path does not affect your application's ability to make LLM calls.
Telemetry retention is configurable per organisation in Account Settings. Telemetry can be disabled for sensitive projects or environments, and data deletion requests are supported.
DoCoreAI activates via a .pth file injected into your
Python environment at install time. If this file fails to load —
due to a misconfigured Python path, version mismatch, or container
build issue — Python silently skips it. Your application continues
running normally, but with zero governance applied.
No error or warning is raised.
A startup health check that explicitly confirms governance is active
is on our development roadmap. Until then, always validate your
installation in a staging environment before production deployment
by confirming DoCoreAI appears in pip list and checking
that your Python environment matches the installation target.
08Compliance & Audit Readiness
DoCoreAI is designed to align with common enterprise security and privacy expectations as we scale.
- SOC 2 Type II — audit targeted Q3 2026
- ISO 27001 control alignment
- GDPR-aligned data minimisation and purpose limitation
- UK GDPR compliant infrastructure (Manchester, England)
- DoCoreAI acts as a data processor — customers remain data controllers
- Subprocessors limited to essential infrastructure providers
- DPA available on request for all paid plans
Building in a regulated industry? See our guide on HIPAA-compliant AI monitoring for how metadata-only telemetry keeps PHI out of scope.
09AI Governance Enablement
DoCoreAI supports post-deployment oversight of GenAI systems without ever capturing prompts or outputs:
- Cost and usage trend monitoring
- Latency and reliability signals
- Retry patterns and behavioural anomalies
- Model and version change visibility over time
This gives governance teams oversight without introducing content-level surveillance.
10Vulnerability Disclosure
Please report it responsibly to [email protected]. We acknowledge all reports within 48 hours and work toward timely remediation. We do not pursue legal action against researchers acting in good faith.
11Contact
Security inquiries: [email protected]
DPA requests: [email protected]
Data residency questions: [email protected]
Privacy information: see our Privacy Policy