Security
Security Whitepaper
This document describes the security architecture we apply to our platforms and to client deployments. It is written for the engineers and risk owners who have to sign off on running our systems, so it stays concrete: mechanisms, boundaries and failure behaviour rather than assurances.
1. Zero-trust access architecture
No request is trusted because of the network it arrives from. Every identity verification, privileged action and administrative call is authenticated cryptographically and authorised against an explicit policy. Long-lived shared secrets are avoided in favour of short-lived credentials, and administrative surfaces are separated from public ones so that a compromised public endpoint does not imply administrative reach.
- Least-privilege roles, reviewed on a fixed cycle and revoked automatically when an engagement ends.
- Tenant identity carried in the token, not inferred from the request body.
- All privileged actions written to an append-only audit log with actor, timestamp and outcome.
2. Sandboxing autonomous workloads
Agentic systems execute code and call tools on their own initiative, which makes containment the central design problem. Generated code runs inside isolated containers or WebAssembly runtimes with no ambient access to the host filesystem or the network; tool access is granted per task through an allow-list, and every call is logged.
- Per-task namespaces prevent lateral movement between concurrent workloads.
- Prompt-injection containment treats all retrieved text as untrusted input, never as instruction.
- Resource limits bound CPU, memory and outbound request volume, so a runaway agent degrades a single job rather than the platform.
3. Data sovereignty and encryption
Client payloads are encrypted in transit with TLS 1.3 and at rest with AES-GCM-256. Records are partitioned per tenant with keys scoped to that partition, so authorisation bugs in application code do not become cross-tenant data exposure. Where a client requires residency in a specific jurisdiction, the deployment is placed there and the boundary is documented.
- Key rotation on a defined schedule, with rotation events logged.
- Post-quantum key exchange for records that must remain confidential for a decade or more.
- Backups encrypted with keys distinct from production keys.
4. Secure development
Changes to production systems are reviewed before merge. Dependency trees are pinned and scanned; build artifacts are produced from a clean environment rather than an operator workstation. Secrets are held in a managed store and injected at runtime — never committed to source control, and rotated immediately if exposure is suspected.
5. Monitoring and incident response
We monitor availability, error rates and authentication anomalies continuously. A suspected incident triggers a defined sequence: contain, preserve evidence, assess scope, notify. Where an incident is likely to result in a high risk to individuals, the competent supervisory authority is notified within 72 hours and affected users are informed without undue delay, in line with the commitments in our Privacy Policy.
6. Reporting a vulnerability
If you believe you have found a security issue in any of our systems, write to dr.jorgeclemente@proton.me with enough detail to reproduce it. We acknowledge reports within two business days, will not pursue legal action against good-faith research that respects user privacy and avoids service disruption, and will credit reporters who wish to be named once a fix is deployed.
Please do not test against live accounts you do not own, and do not access, modify or exfiltrate data that is not yours.
7. Scope and limitations
This whitepaper describes our standard architecture. Individual engagements may add controls required by a client’s regulator or by a data-processing agreement; those additions are documented in the engagement’s own security annex and take precedence for that deployment. This document is not a certification and does not constitute a warranty beyond the commitments in our Terms of Service.