What did U.S. v. Heppner actually rule about Claude?

In February 2026, a US federal court ruled that 31 documents generated through consumer Claude lost attorney-client privilege. The court's reasoning: Anthropic's privacy policy permits data collection and disclosure to third parties, which destroys any reasonable expectation of confidentiality. Anthropic was classified as an electronic communications service provider under FISA 702. Anthropic has acknowledged that Enterprise tiers with Zero Data Retention 'could' lead to a different analysis, but no firm has yet won that argument in court. For European firms, the more direct issue is that the same data still flows to a US entity, which keeps the CLOUD Act and Schrems II problems open.

Why is hardware sealing different from Anthropic's DPA / Zero Data Retention?

A Data Processing Agreement is a contractual promise that a US court can override via the CLOUD Act or a FISA 702 directive, the provider remains technically able to read the prompt. Intel TDX hardware sealing creates an encrypted enclave at the CPU level: memory is decrypted only inside the protected region, and even the cloud operator, the hypervisor, and system administrators are excluded by hardware. Intel DCAP produces a cryptographic attestation per session that proves the workload ran in a sealed enclave. This shifts confidentiality from a contractual promise (overridable) to a technical impossibility (not overridable by a foreign court).

Does Claude for Legal offer EU data residency on the direct API?

Not on the direct Anthropic API. Anthropic offers limited regional options through cloud partners (AWS Bedrock and Google Vertex AI in some EU regions), but these still involve a US controller (Anthropic) as a sub-processor, which keeps the FISA 702 / CLOUD Act exposure intact. Pro, Max and Team consumer tiers do not provide DPA-level EU residency. VoltageGPU is operated by VOLTAGE EI, a French entity (SIREN 943 808 824) in Solaize, with infrastructure in the EU and a native RGPD Article 28 DPA, no cross-border transfer mechanism required.

Can a French law firm be sanctioned by the Bar for using Claude for Legal?

The Conseil National des Barreaux (CNB) reaffirmed in its July 2024 General Assembly report that generative AI use by lawyers must respect professional secrecy as defined by RIN article 2.2 and the law of 31 December 1971. Article 226-13 of the French Penal Code criminalises the disclosure of secret professionnel with up to 1 year of prison and a €15,000 fine. The operative criterion is whether a third party can technically access the contents, not whether they promise not to read them. Using Claude for Legal for privileged work product without hardware sealing, with a US controller subject to FISA 702 and CLOUD Act, creates documented exposure to bar sanctions and potential criminal liability. The Italian Garante already fined OpenAI €15M on 20 December 2024 (provvedimento n° 10085455) for related compliance failures.

How does the pricing compare for a small firm?

Claude Team is publicly priced at $30/seat/month with a 5-seat minimum, so $1,800/year for 5 seats, not including the audit overhead (Schrems II justification per matter, DPIA documentation, US transfer SCC paperwork). VoltageGPU Starter is $349/month for 3 seats ($4,188/year). For a 10-seat firm, Claude Team is $3,600/year vs VoltageGPU Pro at $14,388/year, Claude is cheaper in cash, but does not provide hardware sealing, an EU controller, or a native Art. 28 DPA. For a solo practitioner, VoltageGPU Personal is $20/month vs Claude Pro at $200/month: 90% cheaper with bar-compliant infrastructure.

Claude for Legal Alternative, EU Sovereign Legal AI with Intel TDX Hardware Sealing

Anthropic launched Claude for Legal on 12 May 2026. For European law firms it is structurally unusable on the consumer API.

On 12 May 2026, Anthropic released Claude for Legal with 12 practice-area plugins and over 80 specialized agents covering contract law, employment, litigation, M&A, IP, real estate and regulatory work. The plugins are open-source Markdown and JSON files. The infrastructure is not: prompts and outputs are processed in the United States by Anthropic (San Francisco), with no native EU data residency on the direct API. For European law firms bound by secret professionnel (French CP art. 226-13, up to 1 year of prison and a 15,000 EUR fine), CCBE deontology and GDPR Article 28, this creates a structural problem.

The U.S. v. Heppner ruling (February 2026)

A US federal court held that 31 documents generated through consumer Claude lost attorney-client privilege because Anthropic's privacy policy permits data collection and disclosure to third parties. The court classified Anthropic as an electronic communications service provider under FISA 702. Anthropic acknowledges that Enterprise tiers with Zero Data Retention could lead to a different analysis, but no firm has yet won that argument in court. For European firms, the underlying CLOUD Act and Schrems II exposure remains open regardless of contractual ZDR.

VoltageGPU forked the 12 Claude for Legal playbooks

Anthropic released the playbooks as open-source on 12 May 2026. VoltageGPU forked the 12 plugins and adapted them for French and European law (Code civil, Code du travail, RGPD, CCBE Code de Déontologie, secret professionnel art. 226-13 CP). The playbooks run on Qwen3.5-397B-A17B-TEE and DeepSeek-V3.2-TEE inside Intel TDX hardware enclaves operated by VOLTAGE EI in Solaize, France (SIREN 943 808 824). Memory is AES-encrypted by the CPU, the PCIe link between CPU and GPU is encrypted, and Intel DCAP produces a cryptographic attestation per session. No administrator, including VoltageGPU itself, can read the documents during inference.

Pricing published, no enterprise gatekeeping

VoltageGPU Personal at 20 USD/month for solo practitioners, Starter at 349 USD/month for 3 seats, Pro at 1,199 USD/month for 10 seats, Enterprise at 3,499 USD/month with SSO, SCIM and DeepSeek-V3.2-TEE reasoning. A 5-person team running 12 months pays 4,188 USD on Starter versus 18,000 USD or more on Claude Team for the same headcount, with hardware sealing and a native RGPD Article 28 DPA on top of the 75 percent TCO reduction.

Related comparisons

Claude for Legal Alternative for EU Law Firms, VoltageGPU Sovereign Legal AI

Anthropic launched Claude for Legal on 12 May 2026 with 12 practice-area plugins and over 80 specialized agents covering contract law, employment, litigation, M&A, IP, real estate and regulatory work. The plugins are open-source Markdown and JSON files. The infrastructure is not: prompts and outputs are processed in the United States by Anthropic (San Francisco), with no native EU data residency on the direct API. For European law firms bound by secret professionnel (French CP art. 226-13, 1 year prison and 15,000 EUR fine), CCBE deontology and GDPR Article 28, this creates a structural problem. The February 2026 ruling U.S. v. Heppner held that 31 documents generated through consumer Claude lost attorney-client privilege because Anthropic's privacy policy permits data collection and disclosure to third parties. VoltageGPU forked the 12 open-source playbooks and runs them inside Intel TDX hardware enclaves under a French controller, VOLTAGE EI, on compute operated by listed US sub-processors. Hardware sealing means even the cloud operator cannot read the prompt or the document during computation. Pricing is published: Personal at $20/month for solo practitioners, Starter at $349/month for 3 seats, Pro at $1,199/month for 10 seats, Enterprise custom. A 5-person team running 12 months on Starter pays $4,188 versus Claude Team at $18,000+ for the same headcount, with hardware-rooted confidentiality on top of the 75% TCO reduction.

CLAUDE FOR LEGAL LAUNCHED 12 MAY 2026

Claude for Legal vient de sortir.

Name: VoltageGPU Sovereign Legal AI
Price: 20 USD

Votre Barreau ne vous laissera pas l'utiliser.

12 plugins open-source. 80+ agents juridiques. Hébergé aux États-Unis par Anthropic, sans résidence EU sur l'API directe, et déjà cassé par U.S. v. Heppner (fév. 2026) sur le privilège avocat-client. Nous avons forké les 12 playbooks et les exécutons dans une enclave Intel TDX française.

€20M

RGPD Art. 83, 4 % CA

1 an

Prison, secret pro art. 226-13

Docs Claude sans privilège (Heppner)

Voir l'agent Sovereign Legal Audit GDPR gratuit 15 min

COMPARAISON DIRECTE

Claude for Legal vs VoltageGPU Sovereign Legal AI

Mêmes playbooks open-source (forkés depuis Anthropic). Architecture radicalement différente.

Critère	Claude for Legal	VoltageGPU Sovereign Legal
Controller jurisdiction	United States (Anthropic, San Francisco)	France (VOLTAGE EI, SIREN 943 808 824)
Hardware sealing	No, software promise only	Yes, Intel TDX enclave
Controller and customer database	US controller, US / global processing	French controller, EU-hosted customer database
CLOUD Act exposure	Yes, US subpoena enforceable	Mitigated: French controller, hardware-sealed memory
FISA 702 exposure	Yes, US electronic communications provider	Mitigated: French controller, hardware-sealed memory
Attorney-client privilege	At risk, U.S. v. Heppner (Feb 2026)	Preserved, operator cannot read enclave
GDPR Art. 28 DPA	Enterprise tier only, US transfer via SCCs	Under French law; compute sub-processors listed, transfer documented once
Schrems II re-justification	Required for every EU client matter	Documented once in the DPA (listed US compute sub-processors, TDX-sealed memory)
Training on prompts (default)	Possible on Pro/Max/Team	Never, by design
Retention period	Up to 5 years (consumer tiers)	Zero retention by default
Practice-area playbooks	12 plugins + 80+ agents (open-source)	12 forked playbooks + Sovereign Legal AI agent
Bar deontology fit (FR/EU)	Red flag for secret professionnel	Designed for CCBE / French art. 226-13 CP
Pricing (1 seat / month)	$200-400 (Pro/Team)	$20 Plus, $349 Starter (3 seats)
Live attestation per session	No	Intel DCAP hardware attestation

JURISPRUDENCE, FÉVRIER 2026

Pourquoi U.S. v. Heppner change tout

CE QUE LE JUGE A DIT

31 documents générés via Claude consumer ont perdu la protection du privilège avocat-client. Motivation : la privacy policy d'Anthropic autorise la collecte et le partage avec des tiers, ce qui détruit toute « expectation of confidentiality » raisonnable. Anthropic est qualifié de « electronic communications service provider » au sens du FISA 702.

Anthropic admet qu'un usage « Enterprise » avec Zero Data Retention pourraitdonner une analyse différente. Le mot opératoire est pourrait : aucun cabinet n'a encore obtenu de jugement en sa faveur sur cette base.

POURQUOI L'ENCLAVE TDX CHANGE LA RÉPONSE

Le critère de secret professionnel en droit français (art. 226-13 CP) et CCBE est binaire : un tiers peut-il techniquement accéder au contenu ? Avec une enclave Intel TDX, la mémoire est chiffrée par le CPU, l'opérateur cloud, l'hyperviseur et tout administrateur système sont exclus par hardware, pas par contrat.

Le résultat est une attestation cryptographique par session (Intel DCAP) qui prouve à votre client que sa pièce a été traitée dans une enclave scellée. C'est un fait vérifiable, pas une clause de NDA contournable par un subpoena américain.

TCO, ÉQUIPE DE 5 AVOCATS / 12 MOIS

Le coût caché de Claude for Legal

Claude Team (5 seats)

$18,000+

$30/seat/mo × 5 × 12 mois. Hors audit RGPD, hors SCC, hors DPIA Schrems II requis pour chaque dossier client EU.

VoltageGPU Pro (10 seats)

$14,388

$1,199/mo, 10 sièges, 5,000 requêtes/mois, API CLM. Enclave TDX, contrôleur FR, DPA natif.

Économie + risque évité

$3,612

Sans compter les €20M d'amende RGPD potentielle si l'ordre est informé d'un transfert non encadré.

Pour un cabinet solo : $20/mois sur le tier Plus, contre $200/mois Claude Pro sans garanties EU. 90 % d'économie, et la conformité au Barreau en bonus.

L'ARCHITECTURE QUI CHANGE TOUT

Mêmes plugins. Enclave différente.

Les 12 plugins de Claude for Legal sont publiés en open-source (checklists contrats, redlines, risk scoring, playbooks litigation, M&A, IP, employment). Nous les avons forkéset adaptés au droit français et européen (Code civil, Code du travail, RGPD, CCBE Code de Déontologie, secret professionnel art. 226-13 CP).

Ces playbooks tournent sur Qwen3.5-397B-A17B-TEE et DeepSeek-V3.2-TEE à l'intérieur d'une enclave Intel TDX hébergée en France. La mémoire est chiffrée AES par le CPU. Le GPU est isolé dans le domaine de confiance. L'attestation Intel DCAP est émise à chaque session. Aucune private key ne quitte l'enclave. Aucun administrateur, y compris nous, ne peut lire les documents pendant l'inférence.

C'est la différence entre une promesse contractuelle (DPA, SCC, ZDR) qu'un tribunal américain peut neutraliser via le CLOUD Act, et une impossibilité technique que le hardware Intel garantit indépendamment de toute juridiction.

Vos clients ont droit au secret professionnel.

Audit GDPR gratuit 15 minutes : nous calculons votre exposition Heppner et nous vous montrons comment forker un playbook Anthropic pour votre cabinet.

Voir Sovereign Legal AI Solution cabinets d'avocats