Three changes this week land on systems you rely on but do not run yourself: a cloud contract, a web server, an AI agent on a plant. In each case a contract or a plan settles what happens next.
Items 1 and 2 touch both financial and industrial organisations; item 3 is for industrial ones.
1.Exit plans built on the Data Act may not hold for custom-made cloud contracts signed on or before 12 September 2025
On 21 October the EU member states agreed the Council's position on the Digital Omnibus, a package that simplifies EU data and privacy law. Much of the text amends the Data Act, which gives cloud customers the right to switch provider. Under the Council's text, custom-made cloud services would fall outside the Data Act's switching rules for contracts signed on or before 12 September 2025. Custom-made means most features were adapted for one customer; rented infrastructure such as servers and storage does not qualify. Older contracts with smaller providers, small and medium-sized firms and small mid-caps, get a lighter regime of their own, with the same cut-off and again not for rented infrastructure.
One duty stays for those contracts: switching and egress charges, the fees for moving and for taking your data out, must still come down and end. This is a starting position. Parliament has not voted, and talks between the two have not begun.
If a critical system runs on an older custom-made cloud contract, the rules that make a switch workable may no longer cover it. Those include the provider's duty to help you move and to keep the service running while you leave. For a financial firm, DORA, the EU's digital resilience rules, already expects a tested exit plan for technology services that support critical or important functions. This change makes the contract's own exit terms the part that has to hold. An exit plan that counts on the law can find, halfway through an exit, that only the contract's terms apply.
Ask your team: Which of our cloud contracts are custom-made and signed on or before 12 September 2025, and does each carry its own exit terms?
Sources: Council of the EU, ST 14756/26, mandate for negotiations with the European Parliament (recitals 17–18), 22-10-2026 · Council of the EU, press release on the digital rulebook position, 21-10-2026 · EuropaWire, reprint of the Council press release, 22-10-2026
2.PHP fixed two flaws that let attackers run code on a web server without logging in, and version 8.2 loses support at year end
On 21 October the PHP project released fixes for two critical flaws that let an attacker run commands on a server without logging in. PHP is the programming language behind many corporate websites, web shops and customer portals. The project says every installation that serves web pages is at risk, and that updating is the only safe choice. The same day, the Dutch National Cyber Security Centre rated the flaws high likelihood and high damage. Versions 8.1 and older get no new release; for 7.4, 8.0 and 8.1 the PHP team published bare source patches, and some vendors ship fixes. Support for 8.2 ends on 31 December 2026.
In most organisations a web agency, hosting provider or software vendor runs PHP for you. The risk stays with you all the same. That makes your list of suppliers, and what each one runs for you, the place where this gets settled. A portal left on an unsupported version stays open to this flaw, and after December to every one found later. I wrote four questions to send your web partner.
Ask your team: Which of our websites, portals and supplier-run systems run PHP, on which version, and who has confirmed the update?
Sources: PHP, security advisory "Critical remote code execution vulnerabilities, update now", 21-10-2026 · PHP 8.2.35 release notice (8.2 end of life 31-12-2026), 21-10-2026 · NCSC-NL, NCSC-2026-0443 (PHP), 21-10-2026
3.Schneider Electric's new plant AI agents may start corrective actions on their own
On 21 October Schneider Electric, NVIDIA and HPE announced AI agents for industrial sites that monitor control systems, energy use and cyber security. Schneider says the monitoring agents "recommend or initiate corrective steps", and the security agents "recommend or carry out responses". It says the platform that runs them keeps "human and organisational control of critical operations and infrastructure", without saying how. It gives no availability date.
"Recommend or initiate" is one phrase for two very different risks. For an offshore, marine or manufacturing company, what an agent may change on its own is an operating decision, not a default the supplier picks. The list of actions an agent may take without a person approving belongs in the contract, agreed before the pilot starts. Without it, an agent acting on a default setting can stop a production line or shut down a system on board before anyone agreed it could.
Ask your team: For any AI agent we pilot on plant or vessel systems, which actions may it take without a person approving, and where is that written down?
Sources: Schneider Electric, NVIDIA and HPE press release (EuropaWire), 21-10-2026
From Delfen this week
- On 21 October delfen.com published The PHP emergency patch of 21 October: four questions for your web partner, updated the same day after PHP's release.
- Past issues of this update are at delfen.ai/cto-update.
- Correction to issue 1: I wrote that I found no DORA fines. The Austrian supervisor had already published one on 14 July; issue 1 now carries a correction note.
Checked and quiet
As far as I could find, EN 18286, the European standard for an AI provider's quality system, is still the only harmonised standard cited under the AI Act. The Cbw, the Dutch cyber security act, still has no enforcement action I could find. Under DORA I found no new fines; the only public one I know of is the Austrian supervisor's €42,000 fine, published on 14 July. The first list of standards under the Machinery Regulation, the new EU safety rules for industrial machines that apply from 20 January 2027, is still not out.