In short: a secure legal AI platform proves it, it doesn't promise it. Ask for four things and refuse to move without them. A SOC 2 Type 2 report and ISO 27001 for how the company runs security, ISO 42001 for how it governs the AI itself, GDPR alignment for data subject rights, and AES-256 at rest with TLS 1.3 in transit for the data itself. Everything past that (data residency, no-training guarantees, audit logs, BYOK) is the difference between a chatbot you tolerate and infrastructure your compliance officer signs off on.
Why legal AI security is a harder problem than normal SaaS
Most software handles data. Legal software handles privileged data, the kind that loses its protection the moment it leaks. A marketing tool that spills email addresses is an incident. A legal AI tool that spills a client's litigation strategy can blow attorney-client privilege, and privilege doesn't come back once it's gone.
That raises the bar on two fronts at once. The platform has to be secure in the ordinary sense (encrypted, access-controlled, audited), and it has to be secure in a way that survives legal scrutiny: no third-party training on your matters, hard isolation between clients, and an audit trail you can hand to a regulator. More than 15,000 firms now run legal work through HAQQ, and the ones with in-house counsel or a real compliance function all ask the same questions before they type a single prompt. This is the checklist behind those questions.
The four certifications that actually mean something
Vendors love to list badges. Most of them are marketing. Four are not, because each one is issued after an outside auditor tests the claim, and each one covers a different piece of the problem.
SOC 2 Type 2. This is the one your compliance officer will ask for by name. Type 2 means an independent auditor watched the controls operate over a period of time, not just on the day of the test. It covers security, availability, processing integrity, confidentiality, and privacy of customer data. A vendor that has it can hand you a report. A vendor that says "SOC 2 aligned" but can't produce the report has not been audited.
ISO 27001. The international standard for an information security management system. It certifies that the company has a systematic, documented approach to managing sensitive data, not that one engineer happens to care about security. Your IT department will ask for the certificate. A real vendor hands it over.
ISO 42001. The newest of the four and the one most legal buyers miss. It certifies the AI management system itself, how the model is developed and deployed responsibly, not just how the servers are locked down. For legal AI this matters more than for almost any other category, because the risk isn't only a breach. It's the model doing something you can't govern. If you audit your AI vendors, this is the standard that was written for exactly that.
GDPR. Not a certification, a legal obligation, but the alignment is what proves the vendor treats your data as something a data subject has rights over: access, deletion, portability, and breach notification inside 72 hours. If you handle cross-border matters or anyone's data touches the EU, this is table stakes. Regional privacy regimes like the Gulf's PDPL layer on top of it, so a vendor built for GDPR usually has the machinery to cover those too.
Encryption: what AES-256 and TLS 1.3 actually buy you
Two acronyms come up in every security conversation, and they cover two different moments in the life of your data.
AES-256 at rest. When your documents sit on disk, they're encrypted with a 256-bit key. Strong enough that brute-forcing it is not a realistic threat with current or foreseeable computing. This is the standard governments use for classified data, so it's the floor, not the ceiling, for a legal platform.
TLS 1.3 in transit. When your data moves between your browser and the server, TLS 1.3 (the current version of the protocol behind the padlock in your address bar) encrypts it so nobody sitting on the network can read it in flight. Older versions have known weaknesses. 1.3 is the one to insist on.
The phrase that ties them together is zero-knowledge architecture: the system is built so your data stays private even from the people running it. Encryption at rest and in transit is what makes that claim real instead of a slogan. HAQQ encrypts everything at rest with AES-256 and everything in transit with TLS 1.3, which is the baseline the rest of the controls sit on top of. You can read the full breakdown on the security page.
Where your data lives, and who can touch it
Encryption protects the data. Governance decides who gets to use it, where it sits, and whether it ever leaves your control. Five things belong on the list.
- No training on your data. The single most important line for legal work. Your matters are never used to train the AI, HAQQ's or anyone else's. If a vendor can't say this flatly, assume the opposite.
- Data residency. You choose the region your data is stored in, EU, US, or Middle East, so residency requirements are a setting, not a negotiation. For regulated clients this is often the deal-breaker question.
- Isolation. Each organization's data is logically isolated. No mixing, no cross-client contamination. This is what keeps one firm's matters from ever surfacing in another's.
- Access control. Role-based permissions so people see only what their role needs, with SSO and MFA for enterprise deployments. Privilege leaks are usually access-control failures, not encryption failures.
- Retention and deletion. Configurable retention periods with automatic purging and verification logs, plus full export whenever you want it. You own the data, including the right to make it disappear on your schedule.
For the largest deployments there's one more lever: bring your own key (BYOK). Enterprise customers manage their own encryption keys and control the full key lifecycle and rotation, which means even the platform can't decrypt your data without infrastructure you hold. If that's a requirement, raise it during the enterprise conversation rather than after.
The vendor evaluation checklist
Print this, or paste it into your next vendor call. For each control, ask for the proof in the middle column. If the answer is a shrug, that's your answer.
| Control | Proof to demand | Why it matters for legal |
|---|---|---|
| SOC 2 Type 2 | The audit report, not a badge | Independently verified controls over time |
| ISO 27001 | The certificate | Systematic security management, not one person's habit |
| ISO 42001 | The certificate | Governs the AI model, not just the servers |
| GDPR / privilege | DPA + 72-hour breach terms | Data subject rights and privilege survival |
| Encryption | AES-256 at rest, TLS 1.3 in transit | Data unreadable at rest and in flight |
| No model training | Written, contractual guarantee | Your matters never leak into a model |
| Data residency | Region selection (EU / US / ME) | Meets residency and sovereignty rules |
| Isolation | Logical per-tenant separation | No cross-client contamination |
| Access control | RBAC + SSO + MFA | Privilege leaks are access failures |
| Audit logging | Exportable, complete logs | Compliance evidence and privilege trail |
| BYOK | Customer-managed keys (enterprise) | You hold the keys, not the vendor |
| Incident response | 24/7 monitoring + notification SLA | Breaches get caught and disclosed |
The questions that separate real from theater
Certifications tell you the company passed an audit. These four questions tell you whether the security is designed for legal work specifically.
"Is my data ever used to train a model?" The only acceptable answer is a flat no, in writing. HAQQ's answer is that your data is never used to train its models or any third-party model. Anything softer than that means your matters are potentially feeding a system you don't control.
"How do you handle attorney-client privilege?" A serious vendor has architected for it: complete data isolation so nothing crosses between clients, no attorney data in any training set, and audit trails that satisfy privilege documentation requirements. If privilege is an afterthought bolted on, you'll hear vague reassurance instead of architecture.
"What happens in a breach, and when do I hear about it?" Look for 24/7 monitoring, a documented incident response plan, and notification inside 72 hours per GDPR. A vendor without a clear notification SLA is a vendor hoping you never ask.
"Can I see the report?" The whole test in one line. Certifications that can't be produced don't exist. Third-party audits and penetration tests are only worth something if the vendor can show you the output.
Essayer HAQQ AI gratuitement
Découvrez la rédaction et la recherche juridique par IA
Frequently asked questions
What certifications should secure legal AI have?
Four carry real weight: SOC 2 Type 2 (independently audited security, availability, and confidentiality controls), ISO 27001 (a certified information security management system), ISO 42001 (a certified AI management system, which governs the model itself), and GDPR alignment for data subject rights and breach notification. HAQQ holds all four. Ask any vendor to produce the report or certificate behind each claim, since an unaudited badge means nothing.
Is AES-256 encryption enough to protect legal data?
AES-256 at rest is the right floor, it's the standard used for classified government data, but encryption alone isn't a security program. Pair it with TLS 1.3 for data in transit, logical isolation between clients, role-based access with SSO and MFA, and a written guarantee that your data is never used to train a model. Encryption protects the bytes. Governance decides who can ever touch them.
How do I know a legal AI vendor won't train on my data?
Get it in writing, in the contract, not just on a marketing page. The commitment should cover both the vendor's own models and any third-party models in the stack. HAQQ's position is that client data is never used to train its models or any third-party model, and that you keep complete ownership with export at any time. If a vendor hedges on this, treat it as a no.
Where is my data stored, and can I choose the region?
With a platform built for regulated legal work, yes. HAQQ stores data in SOC 2 compliant data centers and lets you pick your region, EU, US, or Middle East, to meet residency requirements. Regional data residency is increasingly a hard requirement for government and enterprise legal buyers, so confirm it's a setting you control rather than a promise you have to renegotiate later.
Key takeaways
- Four certifications carry real weight: SOC 2 Type 2, ISO 27001, ISO 42001, and GDPR. Demand the report or certificate behind each.
- ISO 42001 governs the AI model itself, not just the servers. Most legal buyers miss it, and it's the one written for exactly this risk.
- AES-256 at rest and TLS 1.3 in transit are the encryption floor. Zero-knowledge architecture is what makes them meaningful.
- The legal-specific controls are no-training guarantees, per-client isolation, regional data residency, and privilege-grade audit trails.
- If a vendor can't produce the report, the certification doesn't exist. "Can I see it?" is the whole test.
Where HAQQ fits
HAQQ was built for legal data from the first line of code, not retrofitted for it. SOC 2 Type 2, ISO 27001, ISO 42001, and GDPR compliance, AES-256 and TLS 1.3 encryption, per-client isolation, regional data residency, no training on your matters, and BYOK for enterprise. The full detail, including the attorney-client privilege architecture, lives on the security page, and the enterprise page covers how large teams deploy it. If your compliance team has a questionnaire, HAQQ has the answers on file.
HAQQ provides legal information and technology, not regulated legal advice. Security and compliance requirements vary by jurisdiction and matter. Confirm your own regulatory obligations with a qualified professional before relying on any vendor's controls.
