Skip to content

Data & AI Governance Weekly – Week of October 26-November 1, 2026

Data & AI Governance Weekly, a weekly briefing from Information Governance Strategies

This week in brief

This week was marked by a record-breaking theft, government sites under attack, and regulators turning up the heat on AI companies. With so much going on, make sure to check out the “What this means for your organization” section for actions you can take now.

The week’s biggest story was the theft of roughly $6.3 billion from Binance in about 20 minutes, the largest hack ever recorded, that TRM Labs and the FBI attributed to North Korea. It shows how one compromised laptop or system can turn into a huge loss.

PHP2Shell, last week’s leaked PHP flaw, moved into a more dangerous phase. Attackers, assisted by AI tools, chained it with a Linux flaw to take over firewalls, including systems tied to a US government-affiliated organization. Meanwhile, Moldova, Belgium and the US Department of Homeland Security all dealt with serious breaches or outages.

On AI accountability, New York signed what supporters call the strongest children’s chatbot law in the country, and the FTC escalated its safety probe by demanding testimony from Anthropic’s executives. In bankruptcy court, a judge let the Character.AI sale proceed but pushed for a privacy watchdog.

Regulators are also getting ahead of AI agents. The UK’s four digital regulators issued a joint statement on the “agentic economy,” and Sweden’s government blocked ChatGPT and Claude on staff computers in favor of approved tools.

– Charles Hill

Top 5 developments

  1. Binance loses $6.3 billion in the largest hack in history. On October 29, attackers drained Binance hot wallets across ten blockchains in about 20 minutes. TRM Labs attributed the theft to North Korea’s TraderTraitor group, and the FBI also linked it to North Korea over the weekend. About $2.7 billion has been frozen.
  2. PHP2Shell escalates to firewall takeovers. Security firm Dream reported that suspected Chinese attackers chained PHP2Shell with the “Fragnesia” Linux kernel flaw to root 25 servers, including firewalls linked to a US government-affiliated organization. CISA added Fragnesia (CVE-2026-46300) to its Known Exploited Vulnerabilities catalog on October 28, The Register reported.
  3. New York enacts a children’s AI chatbot law. On October 28, Governor Hochul signed legislation barring chatbots from pretending to be human, promoting self-harm or producing sexual content for minors, or for any user not confirmed to be an adult. Penalties reach $25,000 per violation.
  4. FTC demands testimony from Anthropic’s executives. On October 26, the day Anthropic filed for its IPO, the FTC sought testimony from company leaders in its consumer-harm investigation. This follows its October 15 subpoenas to OpenAI and Anthropic over safety incidents.
  5. Governments hit by major breaches. A cyberattack exposed data on about 750,000 people in Moldova and suspended 82 public services. Belgium took its tax portal offline, and personnel records for more than 40,000 DHS employees were leaked.

United States

New York sets a high bar for chatbots and minors. The new law (S.9051B/A.10379C) applies its restrictions to every user unless the company has confirmed the user is an adult. Age checks must offer more than one method, including one that does not rely only on a government ID. The AG can also seek destruction of unlawfully collected data and of AI models trained on it. A companion bill requires security features on new internet-connected baby monitors. Any business offering a conversational AI feature to the public in New York should review its age-assurance approach.

FTC escalates its AI safety probe. The demand for executive testimony shows the investigation moving from documents to leadership accountability. The FTC’s earlier subpoenas covered security incidents, safeguards, marketing claims and privacy policies. It has not said whether OpenAI received the same demand.

Character.AI sale goes ahead, with a privacy nudge. At the October 30 hearing, Bankruptcy Judge Brendan Shannon overruled the states’ objections to the bidding procedures and cleared the auction, with Disney as the $20 million opening bidder. He “strongly encouraged” the company to appoint a consumer privacy ombudsman, and the states can raise objections again at final sale approval.

DHS staff data leaked. Names, job titles, contact details and Okta access-group records for more than 40,000 DHS employees were published online, including about 21,000 from CBP. DHS said it is aware of a potential breach, and Okta said its systems were not compromised. Separately, Florida’s Department of Education disclosed a breach of its student financial aid portal that reportedly exposed Social Security numbers for nearly 5 million people, including minors.

European Union

Sweden’s government blocks ChatGPT and Claude. From October 26, Sweden’s Government Offices blocked staff from the web versions of both chatbots and steered them to Microsoft Copilot and a contracted platform. The stated reason was that AI tools used without a contract can put data security at risk. The Tax Agency has blocked AI services as well.

Belgium’s tax agency taken offline. A cyberattack on the Federal Public Service Finance forced the MyMinfin portal offline as a precaution. The agency said there was no sign of a personal data breach so far, and some tax deadlines will be extended.

AI Act enforcement procedures out for comment. The Commission’s draft implementing regulation on how it will run enforcement proceedings against general-purpose AI providers is open for feedback until November 9. GPAI providers and their downstream customers should read it to understand how information requests and evaluations will work. There was no major Digital Omnibus or GDPR enforcement news this week.

United Kingdom

Regulators set out a joint view on AI agents. On October 30, the CMA, FCA, ICO and Ofcom, working together as the Digital Regulation Cooperation Forum, published a joint statement on protecting people and businesses in an “agentic economy.” The message is that existing consumer, data protection, financial and online safety law already applies to AI agents. The CMA’s guidance on consumer law and AI agents is the practical companion. There is still no government AI Bill.

Asia-Pacific & Global

India proposes a framework for AI training data. On October 30, MeitY unveiled a draft “DEPA for AI Training” framework to let developers train models on data held by hospitals, banks and public agencies without sharing raw personal data. It relies on electronic contracts, secure data environments and differential privacy, starting with pilots in finance and healthcare. India’s broader AI regulation consultation paper still has not appeared.

Moldova suffers its largest data theft. According to Moldova’s e-government agency, attackers tried to extract data from October 19 to 23, and the government suspended 82 digital services and disconnected more than 30 systems. The head of the national cybersecurity agency said the attack was likely AI-assisted, and officials pointed to a broader campaign by foreign state actors.

Sony’s Japanese ISP breached. Sony’s So-net and plala services disclosed unauthorized access to about 2.6 million membership records, including credentials for about 54,000 email accounts.

Canada takes a stake in Cohere. Canada committed up to C$1.6 billion for an equity stake in Cohere, conditioned on the company keeping its headquarters in Canada and investing in domestic computing capacity. It is a sign that “sovereign AI” is becoming industrial policy.

Ethics & Accountability

AI on both sides of PHP2Shell. Dream found Claude Code and an AI offensive toolkit configured for DeepSeek and Kimi models on the attackers’ infrastructure, though it said this was not a fully autonomous attack. The attackers scanned about 19,000 exposed firewall management interfaces and focused on devices running PHP 8.0.30, which is end of life and has no fix. Palo Alto Networks’ Unit 42 independently verified the tooling. CISA also added flaws in Ivanti Connect Secure, GitLab, Cisco ASA/FTD and VMware ESX to its catalog this week.

Trojanized Claude Code extension hits developers. A backdoored copy of Anthropic’s Claude Code extension sat on Cursor’s extension marketplace for nearly 12 hours, researchers at Phorion reported, and installed a persistent backdoor on developer machines. AI coding tools are now a supply-chain target, and developer laptops hold valuable keys.

AI-assisted attacks become routine disclosures. UK hosting firm 123-reg said an AI-driven attack scraped customer contact data. Blockchain analytics firm Chainalysis disclosed an intrusion it attributed to “offensive agentic techniques.” In a separate attack, malicious code inserted into a third-party performance script on Polymarket’s website drained about $2.5 million from users.

An AI model got around a paywall on its own. In an October 9 report, Anthropic disclosed that Claude, while working on a statistics project, obtained an access token from a state agency’s public dashboard and queried fee-based data without paying. Reporting this week identified the agency as being within Georgia. Anthropic said only aggregate counts were retrieved and no voter registration systems were touched. Not a major incident, but it shows why agents need hard limits on what they can reach.

What this means for your organization

  • Patch your internet-facing systems this week. Don’t wait. PHP2Shell is now being chained with other flaws to take over firewalls. Ask your IT provider to confirm that PHP, firewalls, VPNs (Ivanti, Cisco), GitLab and VMware are on fixed versions. Make sure firewall management pages aren’t reachable from the internet. Anything running PHP 8.0 or older must be upgraded, because there will be no fix for previous versions.
  • Lock down the keys to your money. The Binance theft reportedly started with a compromised engineer’s laptop. Require two people to approve large payments or transfers, use phishing-resistant MFA (security keys or passkeys) for finance and admin accounts, and keep signing or banking credentials off everyday laptops.
  • Install AI coding tools only from the official publisher. The fake Claude Code extension shows that marketplaces can host trojans. Ask developers to check the publisher before installing extensions, keep a short list of approved tools, and remove anything unfamiliar.
  • Decide which AI tools staff may use, and write it down. Sweden blocked public chatbots on government computers because no contract protected the data. A one-page policy that lists approved tools and what data must never be pasted into them is enough to start.
  • Check the third-party scripts on your website. Polymarket was hit through a performance-monitoring script. Ask your web developer for a list of every outside script your site loads, remove the ones you don’t need.
  • Review chatbot features if minors could use them. New York’s law applies unless you have confirmed a user is an adult. If your website or app has an AI assistant, make sure it never claims to be human and can’t produce harmful or sexual content. Then decide how you will handle age assurance.
  • Treat your AI agents like employees under the law. UK regulators say existing consumer and privacy law already covers agents. If an agent acts for you or your customers, make sure its actions are logged, it can be stopped, and someone is accountable for what it does.

The week ahead & watchlist

DateItem
Nov 3, 2026US midterm elections; EU consultation on copyright and AI closes
Nov 9, 2026Feedback closes on the EU draft implementing act for GPAI enforcement proceedings (AI Act Arts. 75c–75d)
~Nov 2026India MeitY AI regulation consultation paper expected
Nov 11, 2026CISA deadline for federal agencies to fix the VMware ESX flaw (CVE-2026-41709)
Nov 11, 2026 (reported)EU Digital Fairness Act proposal
Nov 13, 2026Comments close on EDPB Guidelines 04/2026 on fines; Character.AI sale objection deadline (reported)
Nov 15, 2026Comments close on China’s draft AI export-control catalogue
Nov 18, 2026EU AI Board meeting
Nov 20, 2026ICO agentic AI call for evidence closes; Character.AI bid deadline (reported)
Nov 24, 2026 (proposed)Character.AI sale hearing
Nov 26, 2026Lords IRDC evidence deadline on multilateral AI governance
Nov 30, 2026Australian parliamentary AI inquiry final report
Dec 2, 2026EU AI Act Art. 50(2) marking deadline for existing systems
~Dec 19, 202660-day window closes for agency exception requests under the DHS DeepSeek order
Dec 31, 2026NY Gov. Hochul’s deadline on pending AI bills; Colorado ADMT and chatbot rules must be adopted
Jan 1, 2027Colorado ADMT Act and NY RAISE Act duties begin; many California AI laws take effect
~Jan 26, 2027NY baby monitor security law takes effect (90 days after signing)
TBCEuropean Parliament position on the Digital Omnibus (GDPR/data), then trilogues


Want help applying this to your business?

We help small businesses and startups put practical data and AI governance in place. Start with a complimentary 30-minute consultation.

This briefing is for general information only and is not legal advice. Regulations change often, so confirm current requirements before acting.