PLEADLY
Security & Compliance
How Pleadly protects attorney-client privilege by design — not by policy.
On February 10, 2026, Judge Jed Rakoff of the U.S. District Court for the Southern District of New York ruled that 31 documents a criminal defendant generated using the consumer version of Claude — and later shared with his attorneys — were not protected by attorney-client privilege or the work product doctrine. Judge Rakoff issued a written opinion on February 17, 2026.
The reason privilege failed was not that the documents were AI-generated. It was that the defendant used a public AI tool whose terms of service allow the provider to collect prompts and outputs, use them for model training, and disclose them to third parties — including government regulatory authorities. Sharing privileged information with a system operating under those terms was treated as a disclosure to a third party, which defeated confidentiality.
This analysis applies to every cloud-hosted AI tool that processes client case content under a standard commercial TOS — regardless of the vendor’s marketing language about “security.”
The court noted that the outcome might have differed had counsel directed the AI use under a Kovel-type arrangement, where the AI functions as an agent of counsel. Pleadly is structured precisely this way: the attorney directs the analysis, the system runs on hardware the firm’s vendor controls under a formal engagement agreement, and no client content reaches any third-party AI provider’s servers.
United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y. Feb. 17, 2026), Dkt. No. 27.
“Does my client’s case content — medical records, accident facts, demand text — pass through a server you don’t control, under terms that allow the vendor to collect, train on, or disclose that data?”
For every cloud-hosted legal AI tool, the honest answer is yes. Pleadly’s answer is no — by architecture, not by promise.
Infrastructure
Authentication, billing, firm settings, case status. Never processes case content.
All AI inference. Document classification, demand drafting, treatment gap analysis. Case content never leaves this plane.
GMKtec EVO-X2 · AMD Ryzen AI Max+ 395 · 96 GB unified memory · Qwen3.5-35B-A3B
Case files, demand drafts, firm memory. LUKS full-disk encryption. No external replication.
The critical boundary: privileged case content — client names, injury details, medical records, police reports, demand letter text — never transits any external API or cloud service. The Intelligence Plane has no inbound access from the public internet.
Compliance
ABA Formal Opinion 512 (July 2024) requires attorneys to understand what their AI tools do with client data — specifically whether the tool uses client information for self-learning, and whether that triggers disclosure obligations. Here is how Pleadly satisfies each relevant requirement.
Case data leaves firm network
Pleadly: No — processed entirely on-premises
Cloud AI: Yes — transmitted to provider cloud
Third-party AI provider access
Pleadly: No — no external AI provider in the pipeline
Cloud AI: Yes — OpenAI / Anthropic / Google
Self-learning on client data (ABA Op. 512 disclosure trigger)
Pleadly: No — on-premises model; no self-learning without explicit written opt-in
Cloud AI: Present — most SaaS legal AI tools require informed consent under Op. 512
Model training on client data
Pleadly: No (opt-in only, anonymized, PII-stripped)
Cloud AI: Varies; often default opt-in
Attorney-directed AI use (Heppner / Kovel)
Pleadly: Yes — attorney directs all analysis through formal engagement
Cloud AI: No — client/paralegal-operated SaaS tool
Encryption at rest
Pleadly: LUKS full-disk encryption
Cloud AI: Provider-managed; varies
Audit log available to firm
Pleadly: Yes — 90-day metadata retention, available on written request
Cloud AI: Rarely
Breach notification
Pleadly: Within 72 hours of confirmed breach
Cloud AI: Varies; often contractual only
Data deletion on termination
Pleadly: Within 14 days of written request, permanent and irreversible
Cloud AI: Varies; often 30–90 days, not always permanent
Pleadly’s ABA 512 Architecture Attestation (v1.3, March 2026) is available for download without a form or email gate. Firms may attach it to their AI vendor due diligence files.
Technical Safeguards
All AI model inference runs on dedicated hardware at Miko Labs’ facility. No case content is transmitted to OpenAI, Anthropic, Google, or any third-party AI provider at any point in the processing pipeline. The language model is served locally via a private inference server with no public internet exposure.
Every request to the intelligence layer is authenticated via HMAC-SHA256 signature. Requests without a valid signature are rejected before any AI processing occurs. Signatures include a timestamp component to prevent replay attacks.
Case data is protected by LUKS full-disk encryption. Encryption keys are never stored on the same physical media as the encrypted data. All data in transit uses TLS 1.3. Internal service communication is isolated to a private Tailscale mesh VPN.
Every claim in a Pleadly demand letter is linked to its source document by ID, page, and character offset. The system requires the AI to return a source_document_id for each citation; outputs without recognized source IDs are rejected before delivery. Hallucinated citations are architecturally prevented, not just warned against.
Each law firm’s case content is stored in a dedicated, org-scoped vector collection. No case content from Firm A is ever in the same query scope as Firm B. Row-level security policies enforce org isolation at the database layer — not just at the application layer.
Pleadly does not use client case data to train, fine-tune, or update AI models without explicit written consent. The attorney review workflow (accepting or rejecting demand drafts) generates anonymized, PII-stripped training signals only when the firm has separately opted in writing. Opting out has no impact on service quality or pricing.
Third Parties
Pleadly uses a limited set of third-party subprocessors for non-privileged functions only. No subprocessor has access to case content, demand letter text, medical records, or any other privileged client information.
Vercel
Control Plane application hosting
Authentication tokens, session management. No case content.
Supabase
Control Plane database
User accounts, firm metadata, billing relationships. No case content.
Stripe
Subscription billing
Payment and subscription data. No case content.
Resend
Transactional email (status notifications)
Notification type and case identifier (no case content, no PHI). Subject lines contain only status codes.
Backblaze B2
Encrypted document backup storage
AES-256-GCM encrypted case document backups only. All files encrypted client-side on Miko Labs infrastructure before upload. Backblaze has no access to encryption keys or plaintext case content.
All AI inference, case data storage, and demand generation occur exclusively on Miko Labs LLC on-premises infrastructure, with no subprocessor involvement.
Legal
The following is excerpted from the Pleadly ABA 512 Architecture Attestation, v1.3, March 2026, executed by Miko Labs LLC:
Responsibilities
Pleadly’s architecture eliminates the primary disclosure risk identified in ABA Opinion 512. However, firms retain responsibility for the following:
All Pleadly-generated demand letters are drafts requiring attorney review before transmission to opposing parties or insurers. Attorneys must exercise independent professional judgment and verify the accuracy of AI-generated content.
Attorneys should evaluate whether disclosure of AI use is required under their engagement agreements or applicable state bar rules. California attorneys should review COPRAC Practical Guidance alongside the Pleadly attestation.
Pleadly does not perform conflict-of-interest checks. Firms must maintain their own conflict-check and matter-opening procedures.
Firms must comply with any state bar ethics rules regarding AI use that differ from ABA Model Rules. Consult ethics counsel for jurisdiction-specific requirements.
Security architecture questions
security@pleadly.aiABA 512 Attestation
Download PDFCompliance review
Book a compliance review call