AI Security for Law Firms: Confidentiality, Retention and Privilege in 2026

AI is already part of legal research, drafting, disclosure review and client communication. The question is no longer whether attorneys will use AI. It is how to use it without compromising confidentiality, privilege or professional responsibility.

This guide is adapted from a briefing BackTier prepared for a small Florida securities-law practice in October 2026. It separates confirmed facts, vendor claims that still need verification, and clearly labeled forecasts. Nothing in it is legal advice.

§ 01The old question and the real question

The old question was "Should we use AI at all?" That is settled. The tools are here, the efficiency gains are real, and courts have already addressed AI use by counsel.

The real question is "How do we use it safely?" Confidentiality, privilege, supervision, disclosure and records preservation do not change because the tool changed. In a securities practice the stakes are higher: material non-public information (MNPI), Regulation FD and SEC disclosure review all demand stricter information controls than most general practice.

Florida Bar Ethics Opinion 24-1 (January 19, 2024) is the controlling ethics authority for Florida lawyers. It permits generative AI — with conditions.

§ 02The news that matters (as of October 2026)

Every item below should be read with one distinction in mind: a vendor announcement is not a contractual guarantee, and neither is an independently verified fact.

• Evidence and discovery. In July 2026 the New York Times and co-plaintiffs moved for sanctions against OpenAI over chat logs allegedly subject to a preservation order. Consumer and Team accounts were covered by that order; Enterprise and Zero Data Retention (ZDR) customers were not. A deleted chat is not necessarily gone.
• Vendor privacy previews. OpenAI previewed "Private Safety Processing" in August 2026. It is a meaningful commitment under scrutiny, not yet independently audited.
• Mandatory retention for frontier models. Anthropic designated its most capable models as "Covered Models" with 30-day mandatory retention, overriding prior ZDR agreements, and later announced Enterprise Frontier Safeguards for customer-controlled storage.
• Gemini as a governable record. Google Vault now covers the Gemini app, so administrators can set retention and litigation holds that override user deletion.
• Legal-specific platforms. Harvey publishes no-training and ZDR commitments in its platform agreement — binding, but only as strong as the contract a firm actually signs.

§ 03AI is becoming infrastructure

Fig. 01 — The same model can be safe or dangerous depending on what it is connected to.

§ 04Six distinctions every attorney must understand

1. 01Training vs. data usage. A vendor can promise not to train on prompts and still retain and review them for safety or compliance.
2. 02Retention vs. deletion. Consumer accounts typically retain inputs for up to 30 days. Deleting a chat removes your view, not necessarily the vendor's copy.
3. 03ZDR is approval-gated. It is not self-serve, does not cover every endpoint, and may not cover frontier models. Confirm eligibility in writing.
4. 04Preservation orders win. Vendor deletion policies are subordinate to court orders. Data that was never retained cannot be preserved.
5. 05Cloud is not air-gapped. No major commercial AI product runs fully air-gapped; private-cloud options are costly and limited.
6. 06Deleted is not destroyed. De-identified logs can still be subject to production. Design the workflow so whatever you delete was never sensitive.

§ 05A simple architecture for a small firm

The architecture controls what the AI can see. That is the entire security model.

• One dedicated AI identity, used only for AI work, never for client email.
• One project per matter, so every session is scoped to a single matter.
• No direct AI connection to client originals — ever.
• Only approved, redacted, non-MNPI files enter the AI workflow.
• A separate cloud backup of approved files, plus encrypted offline drives for required records.
• Any legal-AI tool whose contractual safeguards have not been reviewed stays out of the workflow.

A legal-specific brand name is not automatically safer than a general business platform. What matters is the contract, the architecture and the audit trail.

§ 06What AI can safely do — and what stays protected

Approved workflow (redacted, public or non-MNPI material):

• SEC filing review against published filings and approved redacted drafts
• Regulation FD analysis using public documents
• Press-release and safe-harbor language review
• Legal research and regulatory monitoring (SEC, FINRA, Florida Bar)
• Public educational content: articles, podcast transcripts, client guides

Protected handling required:

• Client-specific MNPI
• Privileged communications and matter strategy
• Unredacted client originals and confidential financial data
• Pre-filing drafts containing non-public information

When in doubt: redact first, then use AI. If it cannot be redacted, it does not enter the AI workflow.

§ 07The legal and ethical reality

• Opinion 24-1. Florida lawyers may use generative AI but must protect confidentiality (Rule 4-1.6), verify output as they would a paralegal's work, bill properly, and make client-facing chatbots identify themselves. Informed client consent is recommended before exposing confidential information to a third-party tool.
• Confidentiality is broader than privilege. A disclosure can survive a privilege challenge and still be an ethics violation. In securities work, Regulation FD applies at the same time.
• Supervision. AI output is not work product until an attorney reviews it. The attorney remains responsible.
• Two false claims to avoid. Using AI does not automatically destroy privilege, and redaction does not automatically make a document safe.

§ 08Forecasts for 2027–2028

These are labeled forecasts, not confirmed events:

• More autonomous legal agents — making matter-level access controls critical.
• Finer-grained enterprise security controls — today's adequate setup becomes tomorrow's minimum.
• Client-imposed AI terms in engagement letters — have a written AI policy before a client asks.
• New model-specific retention rules — review vendor terms when new models launch.
• Audit and provenance requirements — document AI use at the matter level now.
• More AI evidence and discovery disputes — your AI workflow is a potential discovery target.

§ 09Related

• AI Visibility for Law Firms
• AI Dive 046: AI, Security and the Future of Law