Security & Trust
Cordon is built by a certified ISO/IEC 27001 Lead Auditor, so security and compliance are not afterthoughts. They are the product. Here is how we protect your data and hold ourselves accountable.
How we protect your data
Encryption in transit
All data moves over TLS/HTTPS.
Data minimization
Cordon does not retain the raw content of analyzed email by default. We keep the detection results and evidence needed to serve you, and minimize the rest.
Strict tenant isolation
Every record is scoped to your account. One customer can never see another's data, and this is verified by automated security tests.
Password security
Credentials are hashed with Argon2id and never stored in plain text.
Audit logging
Key actions are logged for traceability and audit evidence.
Authentication and access
- Short-lived access tokens with refresh-token reuse detection.
- Least-privilege access to production systems.
- Multi-factor authentication on our administrative accounts.
- Rate limiting and automatic lockout protect our authentication endpoints against brute-force and credential-stuffing attempts.
- Hardened security headers, including strict transport security, are enforced across our services.
- Production systems are configured to expose only what is needed to operate, keeping the surface an outside attacker can reach as small as possible.
- Errors return a generic message to the client. Internal details are never exposed and are captured only in secure internal logs.
How we test ourselves
- Over 800 automated tests run before every release, including a dedicated adversarial suite.
- A daily red-team program attacks our own detection engine with real-world attacker techniques, scored honestly, including what we miss.
- A monthly nation-state / APT “ceiling” test measures our frontier limits.
We publish what we catch and what we miss, because a security vendor that hides its blind spots is a security risk.
AI hardened against manipulation
- Cordon's AI reasons about the intent behind a message, not just surface patterns, as a bounded signal: it can strengthen a verdict but, by design, can never override the deterministic rules or turn a malicious email into a safe one.
- We hardened this AI layer against prompt injection, where an attacker hides instructions inside an email to fool the analyzer. The email is always treated as untrusted data, never as instructions.
- We test it directly: when fed a message instructing the AI to mark itself safe, it refused, and flagged the manipulation attempt itself as a sign of malicious intent.
- Separately, Cordon also detects when an email itself tries to hide instructions aimed at an AI assistant that might read your inbox — invisible text, zero-width characters, encoded payloads — and flags it as a malicious indicator, mapped to MITRE ATT&CK.
Compliance
Every detection auto-maps to ISO 27001, NIST CSF 2.0, SOC 2, and MITRE ATT&CK, and is exportable as audit evidence.
SOC 2 Type II is on our near-term roadmap, as is an independent third-party penetration test. We will update this page as each is completed.
Sub-processors
We rely on a small set of vetted providers: Anthropic (AI analysis), Render (hosting and database), Vercel (web hosting), Mailgun (authorized email), and Cloudflare (network protection). A current list is available on request. See our Privacy Policy for more detail.
Responsible disclosure
Found a security issue? We want to hear from you. Email security@cordoncybersec.us and we will respond promptly. Please give us a reasonable chance to remediate before public disclosure.
Cordon is in early access. This page reflects our current posture and will be kept current as it evolves.