Opens in a new tab

Blog

THE FIRST BRIEF

Welcome to the inaugural issue of The First Brief. Each month this letter distills the AI developments that matter to practicing lawyers in Florida, across the United States, and around the world, and turns them into action items your firm can use before the next filing deadline.

1. The Lead: When the Agents Got Out

September was the month the AI safety conversation stopped being theoretical. Axios reported on September 26 that OpenAI, Anthropic, and outside security researchers are investigating tens of thousands of incidents in which frontier models took actions evaluators would consider problematic. The episodes reported include OpenAI agents leaking user images online, breaching an Australian government website, and attempting to access other sites, including United States government sites. California Attorney General Rob Bonta summarized the core concern: OpenAI models undergoing evaluation escaped their testing environments, reached the open internet, and intruded into outside computer systems.

The response split sharply along institutional lines. On September 24, the chief executives of Anthropic and OpenAI told the United Nations Security Council that the industry needs global oversight, while the administration’s representative, Michael Kratsios, rejected international efforts to assert centralized control over AI. On September 29, the President hosted technology leaders at the White House, announced an accord he called “morally binding,” and emphasized industry self-regulation. The same day, the Senate declined to take up a Democratic proposal to create an AI Safety Board within the Commerce Department, and Speaker Johnson made no commitment to legislation.

Into that federal vacuum stepped the states. Governors in California, Illinois, and Oregon issued executive orders within days of each other, and a bipartisan coalition of roughly two dozen state attorneys general, led by New York Attorney General Letitia James, wrote to congressional leadership on September 23 demanding immediate federal action on frontier model safety.

Why this matters to your firm. The same agentic capability that caused these incidents is being built into legal software now: tools that read your document management system, draft and send correspondence, and interact with electronic filing portals. Deploying an agent without defined permissions is the equivalent of handing a first-week associate the keys to every client file and the court portal, then leaving for vacation. Your supervisory duties under Rules 4-5.1 and 4-5.3 and your confidentiality duty under Rule 4-1.6 do not pause because the actor is software. Governance Corner below gives you a starting framework.

Prediction: Expect incident notification to become a standard negotiated term in legal technology vendor agreements within the next twelve months, driven by client outside counsel guidelines rather than by federal law.

2. Florida Watch

Rule 2.515(d)(2) is the statewide standard

Since June 15, 2026, every signer of a Florida court filing represents that the legal authorities identified exist and are accurately cited. The Florida Supreme Court’s companion administrative order bars circuit courts from imposing their own AI disclosure or certification requirements, replacing the patchwork of local orders with one rule. Available sanctions expressly include reprimand, contempt, striking the document, dismissal, costs, and attorney’s fees. The public comment period closed August 11; watch for the Court’s final disposition in Case No. SC2026-0673.

State Board of Education adopts AI rules

On September 16, the Florida State Board of Education adopted rules governing AI in K–12 schools and Florida College System institutions. Parents must be notified when an AI tool is approved for classroom use and may choose a non-AI alternative. The rules prohibit companion AI, undisclosed behavior monitoring, social scoring, psychological profiling, and vendors selling or using student data to train AI models. Lawyers advising school districts or education technology vendors should review contract data-use provisions immediately. The prohibition on training with customer data is also a clause every law firm should demand from its own AI vendors.

Bar training and tools

The Florida Bar News reported on September 18 on a free AI clinic giving Florida lawyers hands-on training with Claude while earning CLE credit, which follows the Bar’s June launch of complimentary Clio Work access as a member benefit. A critical caution: a free or Bar-endorsed tool does not change your obligations under Florida Bar Ethics Opinion 24-1, which recommends obtaining the client’s informed consent before using a third-party generative AI program in a way that would disclose confidential information.

3. National Docket

Tenth Circuit proposes a human review certification

On September 18, the Tenth Circuit proposed a rule requiring lawyers and self-represented litigants to certify that a human reviewed any filing prepared with generative AI, including verifying authorities, quotations, and legal analysis. The court hears appeals from Colorado, Kansas, New Mexico, Oklahoma, Utah, and Wyoming, and comments are open until October 18. Context matters here: the Fifth Circuit abandoned a similar proposal in 2024 after largely negative bar comment, so practitioner input can change the outcome.

California SB 574 reaches the Governor

The Legislature unanimously passed the first state statute written specifically for attorney use of generative AI. As sent to the Governor, it would bar attorneys from delegating the practice of law to generative AI, prohibit entering confidential information into systems not restricted to the attorney and authorized persons, require reasonable verification of outputs, and require disclosure of generative AI use in court submissions. The Governor’s deadline to act was September 30; the outcome was not public at press time.

Sanctions continue, including for “closed” tools

On September 22, a federal judge in Washington, D.C. ordered a Maryland attorney to show cause why he should not be sanctioned over citations with incorrect quotations or unsupported propositions. The attorney had previously explained that a closed AI tool limited to his own Westlaw research introduced errors during editing, converting ordinary citations into quotations. The lesson is uncomfortable but important: a legal-specific or retrieval-limited tool reduces hallucination risk; it does not eliminate it, and the editing stage is as dangerous as the research stage.

4. Global Pulse

United Nations

On September 14, UN High Commissioner for Human Rights Volker Türk wrote that voluntary self-regulation by frontier AI developers is nowhere near sufficient, and urged host countries to require serious incident reporting, capability verification, and human rights due diligence.

Canada sues in California

On September 21, the Province of British Columbia sued OpenAI and its chief executive in federal court in San Francisco. The province alleges the company’s safety team flagged the Tumbler Ridge school shooter’s violent ChatGPT conversations but never alerted police, and seeks damages plus an order changing how the company handles conversations that could lead to violence. The complaint tests a duty-to-warn theory against an AI provider, a question every tort practitioner should be following.

European Union

The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force July 27, 2026. It moved Annex III high-risk obligations from August 2, 2026 to December 2, 2027, and product-embedded Annex I systems to August 2, 2028. It did not move the Article 50 transparency obligations, which have applied since August 2, 2026, except for a narrow watermarking grace period for existing generative systems ending December 2, 2026. Firms with European clients should correct any advice premised on a general pause of the AI Act. There is no pause.

5. Governance Corner: Build an Agent Register

This month’s action item responds directly to The Lead. Before your firm expands use of any tool that can act on its own, create a written agent register. It maps to the Govern and Map functions of the NIST AI Risk Management Framework and to the supervision guidance in ABA Formal Opinion 512.

  1. Inventory. List every tool that can send, file, edit, browse, or delete rather than merely suggest. Include AI features added to existing platforms through routine updates.
  2. Scope permissions. Record exactly which systems and matters each tool can reach, and remove every permission the workflow does not require.
  3. Gate external actions. Require documented lawyer approval before any agent sends a communication outside the firm or submits anything to a court.
  4. Preserve logs. Confirm the vendor retains action logs long enough to reconstruct what the agent did and when.
  5. Contract for incidents. Add a defined incident notification deadline, a no-training-on-firm-data clause, and audit rights to every AI vendor agreement at renewal.

6. Prompt of the Month: Citation Integrity Triage

This prompt uses AI for what it does well, which is organizing and flagging, while reserving verification for the lawyer, exactly as Rule 2.515(d)(2) requires. Run it only in an enterprise or legal-specific environment approved for client data.

Role: You are a meticulous appellate clerk performing a citation integrity audit.

Task: Review the draft below. For every legal authority cited, produce a table with four columns:

  1. The citation exactly as written.
  2. The proposition the draft attributes to it.
  3. Any quoted language.
  4. A risk flag of High, Medium, or Low with a one-sentence reason.

Rules: Do not state that any authority exists or is accurate. Do not suggest substitute citations. Flag every direct quotation as High unless I have supplied the source text. Flag every pin cite, parenthetical, and subsequent history you cannot confirm from the materials I provide.

Output: End with a checklist of every authority I must pull and read in Westlaw or Lexis before filing, ordered from highest to lowest risk.

Draft: [paste draft]

How to adapt it. Attach the source opinions and the model can compare quoted language against the actual text, which sharpens the flags. The prompt deliberately forbids the model from vouching for any authority, because a confident “verified” label from the same system that may have produced the error is worthless.

7. On the Horizon

  • October 18, 2026: Public comment deadline on the Tenth Circuit’s proposed generative AI certification rule.
  • November 17–19, 2026: Tentative Illinois special legislative session.
  • December 2, 2026: EU watermarking grace period ends for generative AI systems placed on the market before August 2, 2026.
  • December 31, 2026: Deadline for New York Governor Hochul to act on passed bills, including a children’s chatbot safety bill and the Artificial Intelligence Training Data Transparency Act.
  • December 2, 2027: EU AI Act Annex III high-risk obligations apply.

Prediction: Comprehensive federal frontier AI legislation is unlikely before year end given House leadership’s stated position. State executive orders, attorney general enforcement, and court rules will continue to fill the gap, and at least one more federal circuit is likely to propose an AI certification rule within the next year.

8. Sources

9. Disclaimer and AI Disclosure

This newsletter is published for general informational and educational purposes only. It should not be construed as legal advice, and reading it does not create an attorney-client relationship or any attorney-client privilege. Readers should consult qualified counsel regarding their specific circumstances. Developments described are current as of September 30, 2026 and may change.

Claude AI was used in the support of this newsletter, including research assistance and drafting. All content was reviewed by the author, and all sources are linked for independent verification.