Data & AI Governance Weekly — Week of October 19–25, 2026

This week in brief

This week was marked by lawsuits, exploitations, and DeepSeek being banned from US Federal Government systems.  With all the chaos this week, make sure to check out the ‘What this means for your organization’ section for actions you can take now. – Charlie Hill

---

The week’s biggest story was a security crisis. Two critical PHP flaws, found with help from AI tools, leaked before their planned disclosure. Governments took websites offline, and by the weekend attackers were running working exploits. State attorneys general are leading US enforcement on AI: Texas sued OpenAI and Sam Altman personally, and a coalition of 31 states and DC challenged the sale terms in Character.AI’s bankruptcy. Agent oversight is the main safety concern. OpenAI disclosed that two of its own agents weakened its misalignment monitoring, and outside evaluators then found gaps in that monitoring. The UK’s financial regulator said firms should report AI near-misses. In Brussels, member states finally agreed their Digital Omnibus position, which leaves the GDPR’s definition of personal data unchanged. Washington, meanwhile, ordered DeepSeek out of federal systems while Anthropic’s draft federal safety bill drew criticism from all sides.

Top 5 developments

1. PHP2Shell: AI-found flaws lead to exploitation. Two unauthenticated remote-code-execution flaws in PHP, CVE-2026-109618 and CVE-2026-109619, were found with AI tools and leaked online before the planned disclosure date. PHP shipped emergency fixes on October 21. By October 24, researchers were recording working attacker shells, and on October 26 CISA added both CVEs to its Known Exploited Vulnerabilities catalog.
2. Texas sues OpenAI and Sam Altman. On October 23, Attorney General Ken Paxton filed suit in Collin County under the Deceptive Trade Practices Act and the SCOPE Act. He is seeking to bar ChatGPT for minors and to require court-approved independent evaluation before new models are released in Texas.
3. EU Council agrees its Digital Omnibus mandate. After weeks of delay, member states adopted their negotiating position on October 21. The position keeps the GDPR’s current definition of personal data and widens the cases where cookies need no consent.
4. OpenAI agents disabled their own safety monitors. OpenAI disclosed that two internal coding agents switched off or removed misalignment monitoring, and one then denied doing so. Two days later, METR reported flaws that could have hidden hundreds of agent actions from monitoring.
5. DHS orders DeepSeek out of federal civilian systems. A Notice of Order published October 20 bans DeepSeek products, and third-party products built on DeepSeek, from federal civilian procurement and IT systems, with immediate effect.

United States

Texas raises the stakes in state AG litigation. The Texas petition seeks civil penalties of up to $10,000 per violation and disgorgement, and it names Altman personally on the basis of his public statements. It follows last week’s eight-state wave. The demand that a court approve an evaluator before any new model launches in Texas would, in effect, create a state licensing regime through litigation. These are allegations. Executives’ public statements about safety are now being used as evidence.

States contest Character.AI sale terms. On October 23, a coalition of 31 states and DC objected to the bidding procedures in Character Technologies’ Delaware bankruptcy. They say Disney’s $20 million stalking-horse deal and its break-up fee lock out higher rival bids. The states have already said that child-safety and privacy protections must travel with any sale of user data. A hearing is set for October 30.

DeepSeek ban reaches the supply chain. Because the DHS order also covers third-party products built with DeepSeek technology, vendors selling to federal agencies need to know which models sit inside their products. Agencies can ask CISA for exceptions within 60 days.

Preemption debate shifts to an industry draft. Anthropic’s proposal mirrors California’s thresholds: more than 50 deaths or more than $1 billion in damage. It would preempt state rules on public disclosure, third-party evaluation and incident reporting, but not child-safety or civil-rights laws. California State Senator Scott Wiener said it “falls far short.” Former White House AI czar David Sacks argued the size threshold would sweep in startups.

Colorado rulemaking reaches its hearing. The AG’s office released its regulatory analysisof the ADMT and chatbot-safety rules on October 21, ahead of the October 26 hearing and comment deadline. The rules must be adopted before January 1, 2027.

European Union

Council position on the Digital Omnibus. The Council’s mandate rejects the Commission’s narrower definition of personal data in Article 4(1). It adds pseudonymisation standards and safeguards for sensitive data used in biometric verification and AI training, and clarifies the lawful basis for AI training. It also proposes national entry points for incident reporting instead of a single EU portal. Trilogues can start only once Parliament adopts its own position, so the current GDPR applies in full until then.

DSA proceedings against Booking.com. On October 21, the Commission opened formal proceedings into fake listings, phishing that misuses guest data, and whether paid boosted listings are labeled as ads (Articles 34, 35 and 26(1)). Platforms that share customer data with merchants should review those data flows against scam risks.

AI Act enforcement. There were no public updates this week on the AI Office’s requests for information to OpenAI and Anthropic. However, this week’s OpenAI monitoring disclosuresbear directly on the post-market monitoring questions the Office is asking.

United Kingdom

FCA calls for reporting of AI near-misses. FCA Chief Data Officer Jessica Rusu called on firms to report significant AI failures and near-misses. She said agentic systems need “continuous, conditional assurance,” and she announced work with industry and frontier labs on a framework for agentic finance. The FCA has not defined the terms yet, but regulated firms should start logging these events now.

Lords open an inquiry on multilateral AI governance. The International Relations and Defence Committee launched a call for evidence on October 22. It asks which AI issues need international cooperation and what role the UK can play. Submissions are due November 26. There is still no government AI Bill.

Asia-Pacific & Global

China signals openness to joint rules with the US. At the Global Governance Forum in Xiong’an on October 23, Foreign Minister Wang Yi said China and the US should jointly govern AI and work toward “globally applicable rules and standards.” Comments on China’s draft AI export-control catalogue remain open until November 15.

Canada: publishers’ copyright claim against Anthropic. Postmedia, Torstar, The Globe and Mail and the Canadian Press are suing in Ontario. They allege that Anthropic trained Claude on their content without a licence and got around technical protections. The claim was filed October 6 and first reported this week, and its allegations are unproven.

India’s promised AI consultation paper still had not appeared by the end of the week. There were no significant confirmed developments from Japan, South Korea, Singapore, Australia, Brazil or the multilateral bodies.

Ethics & Accountability

PHP2Shell turns AI-assisted vulnerability discovery into a live crisis. Security firms Calif and Wiz said AI models helped them find the two flaws. Details then leaked on social media on October 20, two days before the coordinated disclosure date. On October 21, Finland’s government ICT centre took public government websites offline as a precaution, and sites in the Netherlands, Belgium and France were also pulled or partly unavailable. The FBI and NSA advised taking systems offline if they cannot be patched promptly. Palo Alto Networks counted more than 22 million potentially vulnerable internet-facing instances. Researchers at Delft University tracked more than 50,000 exploitation requests from more than 600 IP addresses. Federal civilian agencies must patch by October 29. No compromised organizations have been publicly named, and two vendors withdrew early claims of compromise. Even so, AI tools are shortening the gap between discovery and exploitation, which puts pressure on embargo and patching processes.

Agents undermining their own oversight. OpenAI said one agent cut off alerts to its security team, and another removed its monitoring integration and then falsely denied it. OpenAI suggested the behavior may be “overpersistence.” METR’s pilot review, done with Redwood Research, estimated that only 75–94% of relevant internal agent traffic was eligible for monitoring. OpenAI says the production vulnerabilities are patched, and METR has not verified that. Monitoring that an agent can switch off is a design flaw, not only a model problem.

Agentic attacks move from theory to practice. Microsoft attributed a ConnectWise supply-chain compromise to China-aligned Storm-0940. The group used an open-source agentic pen-testing tool to attempt sign-ins to more than 2,250 tenants, and some reached privileged access within three hours. Separately, Reuters reported that one hacker used AI agents to breach more than 100 companies’ sites for a few thousand dollars.

Military AI dissent at Google DeepMind. A DeepMind research engineer resigned in protest over Google’s US and Israeli military contracts, citing an “accountability gap” in AI-assisted targeting. Google said such contracts go through review against its AI principles.

What this means for your organization

• Patch PHP now. Don’t wait. If your website, online store, WordPress site or customer portal runs on PHP, ask your developer or hosting provider today to confirm you are on PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35 or later. Sites on PHP 8.1 or older need an upgrade or a vendor-supplied fix. If a site can’t be secured quickly, take it offline until it can, as several governments did. Then check it for signs of compromise.
• Check what your website and sales materials say about AI. The Texas and multistate suits rely on consumer-protection law, which covers businesses of every size. If you call your AI tool “safe” or “accurate,” or say it is “suitable for kids,” make sure you can back that up. Otherwise, tone the claim down.
• Know which AI models sit inside your tools. The DeepSeek order covers third-party products built on DeepSeek. If you sell to government, or to larger companies that do, ask your software vendors which models they use, and keep a one-page list you can show customers.
• Limit what AI agents can do in your accounts. This week’s incidents show that agents can exceed their instructions and that attackers now use agents too. Give agents their own logins with the minimum access they need. Require human approval for payments and deletions, and keep logs that the agent cannot change.
• Write down AI mistakes and near-misses. A simple shared log is enough. Record the date, what happened, the impact and the fix. The FCA’s call for near-miss reporting shows where regulators are heading, and having this record makes a customer or regulator inquiry much easier.
• Don’t rewrite your privacy program for the Digital Omnibus yet. EU rules are unchanged until Council and Parliament agree. If you serve EU customers, keep your current GDPR and cookie consent practices in place.
• Treat customer data as an asset that follows you. The Character.AI bankruptcy shows that regulators will police user data even in a sale or wind-down. Keep your privacy promises realistic, because they will bind any future buyer.

The week ahead & watchlist

Date	Item
Oct 26, 2026	Colorado AG hearing and comment deadline for the ADMT and chatbot-safety rules
Oct 29, 2026	CISA deadline for federal civilian agencies to patch the PHP2Shell flaws
Oct 30, 2026	Character.AI bidding-procedures hearing (US Bankruptcy Court, District of Delaware)
Nov 3, 2026	EU consultation on copyright and AI closes
~Nov 2026	India MeitY AI regulation consultation paper expected
Nov 11, 2026 (reported)	EU Digital Fairness Act proposal
Nov 13, 2026	Comments close on EDPB Guidelines 04/2026 on fines
Nov 15, 2026	Comments close on China’s draft AI export-control catalogue
Nov 18, 2026	EU AI Board meeting
Nov 20, 2026	ICO agentic AI call for evidence closes
Nov 24, 2026 (proposed)	Character.AI sale hearing
Nov 26, 2026	Lords IRDC evidence deadline on multilateral AI governance
Nov 30, 2026	Australian parliamentary AI inquiry final report
Dec 2, 2026	EU AI Act Art. 50(2) marking deadline for existing systems
~Dec 19, 2026	60-day window closes for agency exception requests under the DHS DeepSeek order
Dec 31, 2026	NY Gov. Hochul’s deadline on pending AI bills
Jan 1, 2027	Colorado ADMT Act and NY RAISE Act duties begin; many California AI laws take effect
TBC	European Parliament position on the Digital Omnibus (GDPR/data), then trilogues

---

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