1. Infrastructure
- EU sovereign: all data processed in Amsterdam, Netherlands.
- CPU-only architecture: no GPU dependency for core inspection. Reduces energy consumption by an order of magnitude compared to ML-heavy alternatives.
- Encryption at rest: AES-256-GCM for stored audit reports.
- Encryption in transit: TLS 1.3.
- Deployment: EU-hosted VPS (Amsterdam), single-node. HA topology documented (2 nodes + load balancer + Postgres).
- Semantic inference: external providers with graceful degradation and circuit breakers. EU-based migration in progress.
2. Data handling
- Content handling: prompts and responses are processed to return a verdict. In the current software version the firewall's security log stores the first 10,000 characters of each request and response, plus metadata; a retention job deletes rows after 30 days. Full details: . [DRAFT: a change that stops storing request text by default is in preparation.]
- Metadata: timestamp, endpoint, verdict, latency, source IP (firewall log), keyed hashes (portal).
- Training-data audits: datasets are processed in server memory only and are never written to disk. Using customer content to train models is not part of the current service design. [DRAFT: confirm with counsel.]
3. Access control
- API keys: per-customer, rate-limited, quota-enforced.
- RBAC: internal access restricted to least-privilege roles.
- Audit logging: every API call logged with correlation ID.
- SSRF protection: client-supplied integration endpoints validated before use.
4. Compliance
| Regulation | Status |
|---|---|
| GDPR | Designed to support compliance |
| NIS2 | Designed to support (Belgian law of 26 April 2024) |
| EU AI Act | Designed to support (ATLAS mapping, audit trail) |
5. Security practices
- Security audit before feature development: SQL injection, race conditions, resource leaks, ReDoS, secrets.
- Dependency audit: automated OSV.dev scan on every release.
- No hardcoded secrets: all credentials via environment variables.
- Parameterized SQL only: no f-string or concatenated queries.
- Credential encryption: integration tokens encrypted at rest (AES-128-CBC + HMAC).
- ReDoS protection: Rust regex engine is linear-time; Python regex bounded by separator collapse and analysis window.
- State persistence: trust state and reputation survive restarts. 5-second change-checked flush. Worst-case crash-window: 5 seconds.
6. Incident response
- Breach notification: affected customers within 72 hours (Art. 33 GDPR). [DRAFT: confirm the exact commitment window with counsel.]
- Contact: synaptosec@gmail.com.
7. Responsible disclosure
Scope. Our websites (synaptosec.com and subdomains) and the public API endpoints documented at /docs. Out of scope: physical attacks, social engineering, denial-of-service, attacks on third-party providers, and any testing on other customers' accounts or data.
How to report. Email synaptosec@gmail.com with a description, reproduction steps, and impact. If you have found something sensitive, use encryption when possible. A machine-readable pointer is published at /.well-known/security.txt.
What to expect. We acknowledge reports as soon as we can [DRAFT: response-time commitment to be decided — we only promise what we can keep] and keep you informed of progress. Please give us a reasonable window to fix the issue before any public disclosure; we will credit you in the fix note if you wish.
Safe harbour. We will not pursue legal action against researchers who act in good faith, stay within the scope above, avoid privacy violations and degradation of service, and do not publicly disclose before a fix or an agreed timeline. Belgian law provides conditional protection for good-faith security research; the national route for reporting vulnerabilities and incidents is the Centre for Cybersecurity Belgium (CCB, ccb.bel). [DRAFT: exact legal scope of the Belgian researcher protection to be confirmed by counsel.]
8. MITRE ATLAS mapping
Every blocked event is mapped to a MITRE ATLAS technique, tactic, and URL. Examples:
| Attack | ATLAS Technique |
|---|---|
| Prompt injection | AML.T0051 (LLM Prompt Injection) |
| Tool result injection | AML.T0053 (LLM Plugin Compromise) |
| MCP tool poisoning | AML.T0053 (LLM Plugin Compromise) |
| Memory poisoning | AML.T0051 with propagation flag |
| Data exfiltration via AI | AML.T0025 (Exfiltration via AI) |
| Model extraction / relay | AML.T0044 (Model Extraction) |
| Denial of wallet | AML.T0034 (Cost Harvesting) |
ATLAS mapping is included in every verdict and in every exported security event.