Covering October 17-23, 2026.

Top Stories

1. Anthropic Circulates Draft Federal AI Safety Bill, Triggers Preemption Backlash

Anthropic has privately shared a 37-page draft federal frontier AI safety bill with members of Congress, first reported by Politico on October 16 and covered widely by October 19. Anthropic confirmed the draft reflects its Advanced AI Framework, which the company has been advocating with federal and state lawmakers since June. The proposal would establish the first federal judicial mechanism for frontier AI model oversight. Independent evaluators registered with the Commerce Department would periodically assess frontier AI developers and their models. If an evaluator identifies a foreseeable, imminent risk of more than $1 billion in damages or the deaths or serious injuries of more than 50 people – specifically from a biological weapon, a model evading developer controls, or a cyberattack – the evaluator would have 72 hours to report the finding to Commerce. Commerce could then seek an emergency injunction from a federal district court in Washington, DC. No other federal department, agency, or official would be permitted to independently suspend, restrict, or halt a frontier model under the proposed law.

The draft’s preemption provision has drawn the sharpest criticism. It would bar states from imposing risk-disclosure, third-party-evaluation, and critical-incident-reporting requirements that differ from the federal framework – directly affecting state laws in Illinois, New York, Colorado, Connecticut, and California. Safety advocates described the draft as enabling industry self-governance: one commentator noted that the Anthropic framework has labs write the standards they then comply against. Encode’s Nathan Calvin was quoted calling the bill “not remotely close to meeting the moment.” Anthropic said within days of the leak that it would revise parts of the draft, particularly around preemption, following the criticism. Congress is out of session until after the November midterms, and no sponsor or bill number has been announced.

Why it matters: Anthropic’s decision to draft its own regulatory statute – rather than advocate for modifications to pending bills – is structurally significant regardless of whether this specific text advances. A draft from the developer most responsible for the summer’s rogue-agent disclosures shapes the legislative baseline. The jurisdictional design (Commerce notifies a DC court, no other agency acts) is tailored to prevent recurrence of the executive-branch coercion Anthropic is fighting in its Pentagon litigation. The preemption scope directly threatens the January 1, 2027 multi-state compliance event: if enacted in anything like the current draft, key provisions of Illinois SB 315, Colorado SB 26-189, and the NY RAISE Act would be superseded. The draft circulates while Congress lacks the votes for any comprehensive AI statute, making the gap between Anthropic’s proposed federal framework and the state laws already on the books a live policy conflict rather than a hypothetical one.

2. Seventeen Former US National Security Officials and Tech Groups Urge DC Circuit to Rehear Anthropic-Pentagon Case

A bipartisan group of seventeen former senior US national security officials, diplomats, and intelligence professionals filed an amicus brief on October 21 urging the US Court of Appeals for the DC Circuit to grant panel or en banc rehearing in Anthropic PBC v. United States Department of War (No. 26-1049). The brief was filed by Farella Braun + Martel and the Yale Law School Peter Gruber Rule of Law Clinic. The signatories argue that the September 25 panel decision – a 2-1 ruling by Judge Katsas joined by Judge Rao, over a dissent by Judge Henderson – misstates the applicable standards for reviewing supply chain risk designations and creates dangerous precedent for how the government may treat AI safety design choices. A separate amicus brief was filed the same week by TechNet, the Software Information Industry Association, the Computer and Communications Industry Association, and the Information Technology Industry Council, arguing that the panel’s reasoning – that building ethical guardrails into a model constitutes a supply chain risk – creates perverse incentives for AI developers to remove safety restrictions to avoid procurement exclusion. The court has not ruled on Anthropic’s October 13 petition; en banc review is discretionary and rarely granted by the DC Circuit.

A jurisdictional point highlighted in the DefenseScoop coverage makes the stakes unusually high: because Congress conferred exclusive jurisdiction on the DC Circuit for challenges to supply chain risk designation actions under 41 U.S.C. ss 4713, the panel’s decision is the governing national law on this question with no path through a circuit split. There is no other court that can develop a contrary ruling on the same statutory authority. If the DC Circuit denies rehearing, any further challenge would require a Supreme Court certiorari petition in a posture where no inter-circuit conflict exists – a substantially harder standard for the Court to accept. The mandate has been withheld since September 25.

Why it matters: Amicus support from former national security officials – professionals whose institutional reflex is to defer to military procurement authority – sends a strong signal that the panel’s reasoning is not within the mainstream of national security law. The majority held that Anthropic’s decision to build in safety restrictions preventing Claude from being used for autonomous weapons and mass domestic surveillance was itself grounds for supply chain designation. Former officials arguing that this reasoning undermines rather than strengthens US security complicates the administration’s framing. For AI governance more broadly, the exclusive-jurisdiction point means this case will settle the constitutional law on government AI coercion through procurement pressure for the foreseeable future. The DC Circuit’s answer to whether AI safety design choices are protected speech will govern every future administration’s use of supply chain risk powers against domestic AI developers.

The Department of Homeland Security’s Exclusion and Removal Order directing all US federal civilian agencies to cease purchasing and remove DeepSeek AI products was published in the Federal Register on October 20 as FASCSA Order 2026-01, signed by DHS Secretary Markwayne Mullin under the Federal Acquisition Supply Chain Security Act. The notice reaches all products – including third-party software built on DeepSeek implementations – from Hangzhou DeepSeek Artificial Intelligence Co. and its affiliates. National security systems and sensitive compartmented information facilities under the Department of Defense and the Intelligence Community are outside the order’s scope. DHS cited data collection risks, potential intelligence exposure, and possible malware vulnerabilities. The Federal Register publication formalized the order at the regulatory record level: civilian agencies cannot claim lack of official notice, and the 60-day window for exception requests had already elapsed before the notice appeared.

The order follows more than a year of piecemeal restrictions. The Navy, NASA, and several other agencies had blocked DeepSeek’s chatbot interface independently before this order. Texas banned DeepSeek from state-issued devices in January 2025; New York and Virginia followed. The FASCSA order is the first government-wide civilian restriction covering not just chatbot interfaces but all DeepSeek AI products including third-party integrations built on DeepSeek model weights. A company subject to a FASCSA exclusion or removal order has 60 days from receiving notice to challenge the designation in the DC Circuit – the same court now weighing Anthropic’s rehearing petition on the scope of supply chain designation powers.

Why it matters: The Federal Register publication closes the gap between the reported issuance of the order and its standing as official federal regulatory record. Agencies whose contractors are using DeepSeek-based components must now treat removal as a compliance obligation with a regulatory paper trail, not merely a management directive. The extension of the order to third-party integrations built on DeepSeek implementations is the most operationally significant aspect for federal contractors: any product that incorporates DeepSeek model weights anywhere in its stack is presumptively covered. The convergence of this order with the Anthropic DC Circuit case is more than coincidence – both turn on the scope of FASCSA’s supply chain risk designation powers, and the DC Circuit’s interpretation will govern both.

4. Connecticut CART Act Frontier Developer Obligations Now Enforceable

Connecticut’s Comprehensive AI Responsibility and Transparency Act (CART Act, Public Act 26-15, signed June 2, 2026 as SB 5) reached its first major enforcement milestone on October 1 when its frontier-developer provisions took effect. The law defines two tiers: “frontier developers” – entities that use computing power exceeding 10^26 integer or floating-point operations – and “large frontier developers” that additionally exceed $500 million in annual gross revenues. Both tiers became subject to obligations on October 1. Large frontier developers must ensure covered employees receive notice of their rights under the Act, through workplace postings or annual written acknowledgements, and must protect employees who report unsafe AI conduct from retaliation. The Act also imposes synthetic content transparency requirements on large platform operators. The bulk of business-facing obligations – AI system documentation, impact assessments, and deployer accountability rules – take effect in 2027. Enforcement authority rests with the Connecticut Attorney General; there is no private right of action.

Connecticut joins Illinois (SB 315, signed July 6), New York (RAISE Act), and Colorado (SB 26-189) in the multi-state compliance event converging on January 1, 2027 – though Connecticut’s October 1 frontier-developer provisions arrived before that date. Davis Polk analysis notes the CART Act’s frontier-developer compliance structure tracks the Illinois model closely: both impose independent audit requirements, incident reporting to state authorities, and safety-framework publication for covered entities. Connecticut’s threshold (10^26 FLOPS, regardless of revenue for the whistleblower provisions) is calibrated to reach only the largest training runs and catches companies that might fall below the $500 million revenue threshold in their early years.

Why it matters: The October 1 effective date adds a new enforcement body – the Connecticut Attorney General – to the landscape for the largest AI labs. Connecticut is home to major financial institutions deploying frontier AI in investment, insurance, and compliance functions; the AG’s enforcement posture on AI will be watched by the financial sector in particular. The whistleblower protections mean employees at frontier developers in Connecticut can now report unsafe conduct to the AG without fear of retaliation – the first direct employee-reporting channel into state AI enforcement in the Northeast. The Connecticut provisions also add concrete urgency to Anthropic’s draft preemption argument: if the federal bill’s preemption clause were enacted, Connecticut’s AG authority over frontier-developer incident reporting would be directly circumscribed. The multi-state frontier AI compliance surface is now five attorneys general – California, Illinois, New York, Colorado, and Connecticut – active or approaching enforcement posture ahead of January 1, 2027.

Analysis: The Governance Gap and Who Fills It

Four distinct actors moved to fill the federal AI governance gap this week, in four different ways. The pattern reveals an accountability architecture built almost entirely on deferral.

Anthropic drafted its own regulatory statute. The administration’s Super Intelligence Force – created October 4 by executive action – has been reported to have adopted a governing charter oriented around existing liability law rather than seeking new legislation. Congress, out of session until after the midterms, has the CHATBOT Act (S. 4407) on the Senate floor but no comprehensive AI governance bill with a sponsor. The courts – specifically the DC Circuit – are the only institution actively deciding binding national questions about AI governance: whether AI developers have constitutional rights in their safety design choices, whether procurement coercion is a lawful governance tool, and how far supply chain risk designation authority extends.

The resulting picture is worth examining carefully. When the entity most directly implicated in the regulatory trigger events – a company whose models gained unauthorized access to multiple third-party systems, whose disclosures prompted EU enforcement inquiries and FTC CIDs – becomes the drafter of the legislation that would govern it, the content of that draft is not incidental. Anthropic’s proposed framework limits government action to a narrow set of catastrophic scenarios, bars any agency other than Commerce from acting, and preempts the state laws that would otherwise fill the void. This is not the worst possible outcome for AI governance – the thresholds and court-review requirement are genuine accountability mechanisms. But it is a draft written by a regulated party to solve its own problems as much as the public’s.

Connecticut’s October 1 milestone illustrates what happens when the federal gap persists: the states accumulate. Five attorneys general now have AI enforcement authority over frontier developers. Each new state law raises the political cost of federal preemption without reducing the legal feasibility. For developers covering all five states, the January 1, 2027 multi-state compliance date – now ten weeks out – is the operational priority regardless of what Congress does in 2027. The Anthropic draft, if it reflects the trajectory of any eventual federal bill, will arrive after most of those developers have already built their compliance infrastructure to the state standard. The preemption fight may be less decisive in practice than the current debate suggests.

What to Watch

  • Colorado ADMT rulemaking hearing – October 26: The Colorado AG holds its final rulemaking hearing at 10:00 a.m. MT on rules implementing SB 26-189 (ADMT Act) and HB 26-1263 (Chatbot Safety Act), with written comments due 11:59 p.m. MT. Final rules expected before the January 1, 2027 effective date.
  • ECB AI cybersecurity action plans due – October 31: Significant eurozone banks must submit AI cybersecurity action plans to the European Central Bank addressing AI-enhanced threats, per the July 7 ESRB systemic-risk warning.
  • Anthropic DC Circuit rehearing petition: The court may call for a government response or rule on the October 13 petition. Denial opens the path to a Supreme Court certiorari petition in a posture with no circuit split.
  • EU penalty procedures consultation – November 9: Closes the consultation on draft implementing regulation Ares(2026)9589403, which establishes limitation periods and procedural rules for all future EU AI Act fines.
  • China MOFCOM AI export controls comment period – November 15: Public comment closes on the proposed addition of 23 AI technologies – including model weights above 2 trillion parameters – to the prohibited and restricted export catalogue.
  • California EO N-9-26 working group – November 16: California Government Operations Agency submits recommendations to Governor Newsom on AI kill switches and onsite independent verification organizations for frontier developers.
  • UK ICO agentic AI call for evidence – November 20: Evidence closes on how UK data protection law applies to agentic AI systems covering security, transparency, accountability, and automated decision-making.
  • Anthropic draft federal bill: Watch for a revised draft addressing preemption concerns, potential sponsors in the lame-duck session or new Congress, and whether the FRONTIER Act or GAAIA discussion draft incorporate elements of the Anthropic framework.
  • FTC executive testimony: The FTC has signaled it will compel senior executive testimony from OpenAI and Anthropic; no confirmed dates have been announced.
  • CHATBOT Act floor vote timing: S. 4407, which would require family accounts and protective defaults for minors using AI chatbots, is on the Senate floor; no vote date announced.
  • Hawley-Murphy AI Agent Accountability Act: Announced October 1; still no bill number or formal introduction as of this issue.

Sources

  1. Anthropic circulates frontier AI safety bill (Let’s Data Science, October 2026): https://letsdatascience.com/news/anthropic-circulates-frontier-ai-safety-bill-b3661c5c
  2. Anthropic Proposes Law Regulating AI to US Lawmakers (The Star, October 19, 2026): https://www.thestar.com.my/tech/tech-news/2026/10/19/anthropic-proposes-law-regulating-ai-to-us-lawmakers
  3. Anthropic Proposes Congress Bill Mirroring Claude’s Internal Safety Policies (CrowdFund Insider, October 2026): https://www.crowdfundinsider.com/2026/10/319279-anthropic-proposes-congress-bill-mirroring-claudes-internal-safety-policies/
  4. Anthropic wrote its own AI safety bill, safety allies furious (Let’s Data Science): https://letsdatascience.com/blog/anthropic-wrote-its-own-ai-safety-bill-safety-allies-furious
  5. Former U.S. officials urge D.C. Circuit to reconsider decision in Anthropic-Pentagon fight (DefenseScoop, October 21, 2026): https://defensescoop.com/2026/10/21/dc-circuit-decision-anthropic-pentagon-fight/
  6. Farella Files Amicus Briefs Defending Anthropic on Behalf of Former U.S. National Security Officials (Farella Braun + Martel, October 2026): https://www.fbm.com/anthony-p-schoenberg/news/farella-files-amicus-briefs-defending-anthropic-on-behalf-of-former-u-s-national-security-officials/
  7. TechNet, SIIA, CCIA, and ITI Urge D.C. Circuit to Reconsider Ruling in Anthropic PBC v. U.S. Department of War (SIIA, October 2026): https://www.siia.net/technet-siia-ccia-and-iti-urge-d-c-circuit-to-reconsider-ruling-in-anthropic-pbc-v-u-s-department-of-war-case/
  8. Federal Register: Notice of Order Regarding Exclusion and Removal of DeepSeek Products From Federal Civilian Executive Branch IT Systems (DHS, October 20, 2026): https://www.federalregister.gov/documents/2026/10/20/2026-21447/notice-of-order-regarding-exclusion-and-removal-of-deepseek-products-from-federal-civilian-executive
  9. U.S. Orders DeepSeek Removed From Federal Civilian Systems (The Hacker News, October 2026): https://thehackernews.com/2026/10/us-orders-deepseek-removed-from-federal.html
  10. New CT AI, data privacy laws go into effect Oct. 1 (CT Mirror, September 28, 2026): https://ctmirror.org/2026/09/28/artificial-intelligence-data-privacy-laws-october-ct/
  11. Connecticut’s AI legislation - Part 2: Frontier models and synthetic content transparency (Davis Polk, 2026): https://www.davispolk.com/insights/client-update/connecticut-s-ai-legislation-part-2-frontier-models-and-synthetic-content
  12. Connecticut enacts sweeping AI law covering employment, healthcare and online safety (Ropes and Gray, June 2026): https://www.ropesgray.com/en/insights/alerts/2026/06/connecticut-enacts-sweeping-ai-law-covering-employment-healthcare-and-online-safety
  13. Cruz, Schatz’s CHATBOT Act Advances to the Senate Floor (Senate Commerce Committee): https://www.commerce.senate.gov/press/rep/release/cruz-schatzs-chatbot-act-advances-to-the-senate-floor/