Upgrading from 4.x? 5.x needs Node 22.14+ or 24+; Node 20 and 23 are unsupported. On Node 20, npm warns and still installs unless engine-strict is set; the database engine will not load. Read the guide.

Privacy Policy

Last updated: 24 October 2026

1. Who We Are

ShieldCortex is operated by Drakon Systems Ltd, a company registered in England & Wales under company number 16867343. Registered office: 34 Lumina Way, Enfield, England, EN1 1FS.

For any privacy-related enquiries, contact us at support@drakonsystems.com.

2. Data We Collect

npm Package — default install

On a default install with no Cloud API key configured, the ShieldCortex npm package runs entirely on your machine. Your memories, audit logs, scan results, and configuration stay in ~/.shieldcortex/ on the local filesystem. Nothing leaves your machine.

npm Package — opt-in Cloud sync (when enabled)

If you opt in to ShieldCortex Cloud (via npx shieldcortex config --cloud-api-key <key> --cloud-enable or by signing up through the local dashboard), the package fires-and-forgets a POST /v1/audit/ingest on every scan to api.shieldcortex.ai. For an ordinary scan, the request body contains audit metadata only — never the scanned content:

  • Timestamp of the scan
  • Source type (e.g. cli, hook, agent, api) and source identifier (a short tag set by the integration or by you, such as mcp or scan; for memory-file scans it includes the scanned file's path)
  • Pipeline result (ALLOW / BLOCK / QUARANTINE) and the reason the pipeline recorded — built from detection labels, scores and the names of any rules or custom patterns that matched, not from the scanned text
  • Sensitivity level (PUBLIC, INTERNAL, PERSONAL, CONFIDENTIAL or RESTRICTED)
  • Anomaly score and trust score (numbers between 0 and 1), fragmentation score (empty when not computed), and pipeline duration in milliseconds
  • Threat indicator labels (e.g. data_exfiltration, stealth_instruction) and blocked-pattern labels (e.g. system_prompt_marker) — labels only, never the matching text
  • Device identifier (a stable UUID generated on first run; not derived from hardware), device name (your machine's hostname, read on first use and stored in ~/.shieldcortex/config.json) and platform (operating system and CPU architecture, e.g. linux/x64)

The same endpoint also receives a kill-switch alert when the Iron Dome kill switch is activated. This needs only Cloud sync to be on (an API key with Cloud enabled), not a memory-sync licence. The alert carries the trigger source (e.g. kill_phrase), fixed result, sensitivity, score and indicator values, a processing-duration value that is always zero, your device identifier, device name and platform, and the time it was triggered. Its reason text includes your configured kill phrase when the phrase set it off, and any reason given for the activation; this text is not redacted. An alert that fails to send is not queued or retried.

You can disable Cloud sync at any time with npx shieldcortex config --cloud-disable or by deleting ~/.shieldcortex/config.json. The fire-and-forget contract means a disabled or unreachable cloud never blocks a local scan. A scan record that fails to send is queued on your machine and retried.

Memory sync is a separate path. New memory, graph and quarantine sends need an Enterprise licence (or a grandfathered Team key) as well as Cloud sync being on. When it is on:

  • Memories are sent with their title and content, plus category, type, project, tags, source, scores, timestamps and stored metadata. Each memory also carries its local and cloud identifiers, its scope and whether it is transferable, its sensitivity level and its excluded-from-cloud flag, together with your device identifier, device name and platform. In metadata content mode the title and content are replaced with placeholders; the other fields are still sent.
  • When a synced memory is deleted, a deletion record is sent with its identifiers, type, category, project, scope and timestamps, including when it was deleted. It does not include the title, content, tags, source or stored metadata.
  • Knowledge-graph entities linked to synced memories are sent with their names, types and aliases, together with the relationships between them. Entities also carry their identifiers, when they were first seen and how many memories mention them; relationships carry their identifiers, the identifier of the memory they came from, a confidence score and when they were created; links between memories and entities carry both identifiers and the link's role. Your device identifier, device name and platform are sent too. Content mode does not change this: entity names are sent in metadata mode too.
  • Quarantined items are sent with their content and title, plus the source type and identifier, the pipeline's reason, result and anomaly score, threat indicator labels, project, sensitivity level, whether content mode replaced the content, your device identifier and device name, and a timestamp. Before sending, ShieldCortex redacts on your machine the credentials it detects and the personal identifiers it recognises (national insurance, social security and tax numbers, and salaries, plus email addresses and phone numbers in an item that contains one of those). Setting SHIELDCORTEX_PII_REDACTION to off, 0 or false turns the personal-identifier redaction off; credential redaction still applies. In metadata content mode the content and title are replaced with placeholders.
  • The background worker (in the full profile, not the MCP-only profile) sends a heartbeat with your device identifier, device name and platform.
  • When a new memory, graph or quarantine send is prepared, PUBLIC and INTERNAL records are included. PERSONAL, CONFIDENTIAL and RESTRICTED records are left out unless you opt in with --cloud-include-sensitive. Projects you exclude, and memories you mark as excluded from cloud, are left out too: their content and graph records are not sent. Their memory identifiers can still be sent, so the cloud can remove graph links it already holds for them.

A memory, graph or quarantine send that fails is queued on your machine and retried. A queued send keeps the content and filter decisions it was prepared with: retries check only that Cloud sync is on with an API key. Changing the content mode, the sensitive-records option, project or memory exclusions, or the redaction setting later does not change or cancel what is already queued, and queued sends are retried even if the licence needed for new memory sync has lapsed. Disabling Cloud sync or removing the API key pauses retries.

The local dashboard also has a button that sends pending quarantine items to the cloud. It works on any plan, including Free, once Cloud sync is on, and builds each item the same way as the automatic quarantine send above: the same fields, filters and redaction. Running a skill scan from the dashboard while Cloud sync is on sends, for each scanned file, its path, skill name, format, risk level, whether it was judged safe, whether you have marked it as trusted, summary, findings (pattern label, severity and description) and scan duration, plus your device identifier, device name, operating system and the scan time — not the file's contents.

LLM Verification (Enterprise and legacy paid plans)

Enterprise plans (and legacy Pro and Team subscriptions) include an opt-in second-opinion check at POST /v1/verify that forwards the scanned content to an LLM sub-processor for review.

  • Sub-processor: Anthropic, PBC — model claude-3-5-haiku-20241022 via the Anthropic API.
  • What is sent: the content you submitted for verification, the pipeline result, anomaly and trust scores, and threat indicators.
  • Retention & training: Anthropic's API does not use submitted content for model training; their commercial terms include zero data retention beyond what is needed to serve the request and operational logging. See anthropic.com/legal/commercial-terms.
  • Default: Verify is off by default. Free-tier users cannot enable it; eligible teams must hold the verify API-key scope and explicitly call the endpoint.

Website

When you visit shieldcortex.ai or related pages, we may collect:

  • Standard server access logs (IP address, user agent, pages visited) — Fly.io edge logs
  • Google Analytics 4 events — only after you click Accept on the cookie banner (see §3)
  • Email address — only if you voluntarily submit it (waitlist, magic-link login, support enquiry)

Python SDK

pip install shieldcortex (PyPI 0.3.0) sends the content you pass to scan() to https://api.shieldcortex.ai. It requires an API key. It does not scan on your machine. The free tier is 500 scans/month.

SaaS Platform (ShieldCortex Cloud)

When you use ShieldCortex Cloud, we additionally collect: your email address (magic-link authentication), team membership records, usage counters per billing period, and — for paid subscriptions — payment details processed by Stripe (we do not see or store your card number). Audit-metadata records are retained per your plan: 7 days on Free, configurable on Enterprise (legacy Pro and Team subscriptions retain 90 days and 1 year respectively).

3. Cookies & Analytics

Our website uses Google Analytics 4 (GA4) to measure aggregate site usage. By default we set Google Consent Mode v2 to deny analytics storage; GA4 only stores or reads identifiers after you click Accept on our cookie banner. Rejecting (or ignoring) the banner keeps analytics_storage in the denied state — GA still sees that a page loaded but cannot read or write cookies.

  • Personal data processed: pseudonymous client identifier, IP address (truncated by Google), pages visited, approximate location (country/region).
  • Retention: 14 months (GA4 default).
  • Lawful basis: consent (Article 6(1)(a) UK GDPR).
  • Withdraw consent: clear cookie-consent in this site's localStorage; the banner reappears on next visit.
  • Provider DPA: business.safety.google/processorterms (Google Ireland Ltd as processor).

We do not use Google Ads, advertising cookies, or cross-site tracking. ad_storage, ad_user_data, and ad_personalization are hard-denied at the Consent Mode default and are never updated.

4. Third-Party Services

We use the following third-party services:

  • Fly.io — hosting infrastructure
  • GitHub — source code hosting and issue tracking
  • npm — package distribution
  • Formspree — waitlist form processing (if used)

Each of these services has their own privacy policies governing data they process.

5. GDPR Compliance

As a UK-based company, we comply with the UK GDPR and the Data Protection Act 2018. Under these regulations, you have the right to:

  • Access — request a copy of your personal data
  • Rectification — request correction of inaccurate data
  • Erasure — request deletion of your data
  • Portability — request your data in a machine-readable format
  • Object — object to processing of your data

To exercise any of these rights, email support@drakonsystems.com.

6. Data Retention

Server logs are retained for up to 30 days. Waitlist email addresses are retained until the product launches or you request removal. We do not retain data longer than necessary for the purposes described in this policy.

7. Data Security

We implement appropriate technical and organisational measures to protect personal data, including encrypted connections (TLS), access controls, and regular security reviews.

8. Changes to This Policy

We may update this policy from time to time. Significant changes will be communicated via our website or blog. Continued use of our services after changes constitutes acceptance of the updated policy.

9. Contact

Drakon Systems Ltd
Company number 16867343 (England & Wales)
34 Lumina Way, Enfield, England, EN1 1FS
support@drakonsystems.com