Brussels Tweaks the GDPR Playbook—But Nothing Is Law Yet
On paper, Europe is trying to clear some regulatory underbrush so homegrown AI builders can train on European data without drowning in consent pop-ups. On October 21, 2026, ambassadors of EU member states in the Permanent Representatives Committee agreed on the Council of the European Union's negotiating mandate for the Digital Omnibus package (procedure 2025/0360(COD), Council document ST 14756/26), following up on the European Commission's November 19, 2025 proposal. One centerpiece of the draft text addresses how companies can invoke legitimate interest under Article 6 of the General Data Protection Regulation (GDPR) to train and operate AI systems.
Legitimate interest is not a free-for-all. Back on December 17, 2024, the European Data Protection Board (EDPB) adopted Opinion 28/2024, confirming that legitimate interest can support AI model development and deployment—such as conversational assistants or cybersecurity tools—only if controllers pass a strict three-step necessity and rights-balancing test. The Council's mandate codifies that principle in a proposed Article 88 bis, clarifying that building an AI system is not automatically a legitimate interest on its own and remains subject to case-by-case balancing. Crucially, the Council dropped the Commission's earlier idea of an unconditional right to object in favor of the standard Article 21 GDPR mechanism, meaning objectors can force a halt unless the developer proves compelling legitimate grounds that override individual rights, or needs the data for legal claims. Even then, engineers face the messy reality of machine unlearning: scrubbing an objector's data before training is straightforward if you know who they are, whereas provably erasing their statistical footprint from an already-trained frontier model without a full retrain remains an open research headache.
For IT and privacy teams, there is one critical caveat before anyone rips out a single compliance control: the GDPR Digital Omnibus is not law yet. The Council's October 2026 text is merely its opening bid for interinstitutional negotiations (trilogues) with the European Parliament. Parliament's joint Industry (ITRE) and Civil Liberties (LIBE) committees are still wading through more than 1,750 amendments to their June 2026 draft report, pushing a likely committee vote into spring 2027 and practical application potentially into 2028. (By contrast, a separate track—the Digital Omnibus on AI, Regulation (EU) 2026/1744—already took effect on July 27, 2026, deferring high-risk AI Act deadlines.) In short, existing GDPR obligations, 72-hour breach reporting rules, and ePrivacy cookie requirements remain fully in force today.
The Frontier Compute Gap and Europe's Catch-Up Bid
Even if Brussels streamlines data access tomorrow, turning raw records into frontier intelligence requires staggering quantities of electricity, capital, and accelerators. According to research tracked by Epoch AI, training compute for cutting-edge language models has grown roughly fivefold per year since 2020, while training costs have climbed about 2.4x annually since 2016. At the same time, algorithmic efficiency has roughly halved the compute needed to match a fixed benchmark every eight months between 2012 and 2023. Put those two curves side by side and you get an asymmetric race: catching up to yesterday's frontier gets cheaper over time, while pushing the new frontier gets exponentially more expensive.
Stanford HAI's 2026 AI Index Report illustrates how concentrated the top tier remains. Drawing on Epoch AI data, the report attributes 59 notable model releases in 2025 to the United States and 35 to China, compared with just two in Europe—one from France and one from the United Kingdom. (As Stanford's methodology notes, national attribution reflects institutional affiliations of paper authors rather than a pure measure of domestic industrial capacity, and multi-country collaborations can be counted in more than one column.) Performance gaps can also swing quickly: Stanford's report found that in March 2026, Anthropic's leading model edged out its nearest Chinese rival by just 2.7% in Arena head-to-head user preference rankings, while Chinese open-weight model DeepSeek-R1 briefly matched the top US model in February 2025—though closed models widened their lead over open models from 0.5% in August 2024 to 3.3% by March 2026.
Europe is trying to buy its way onto that playing field, albeit on a slower clock than Silicon Valley hyperscalers:
- The EUROPA Consortium: On June 19, 2026, the European Commission and the EuroHPC Joint Undertaking selected the Italian-led EUROPA consortium (headed by Domyn) as the winner of the Frontier AI Grand Challenge. The project aims to build a fully open-source frontier model of at least 400 billion parameters covering all 24 official EU languages, backed by up to 2.5% of EuroHPC's supercomputing capacity for a year. As of November 2026, however, that model had not yet shipped.
- AI Gigafactories: On July 30, 2026, EuroHPC opened tender
EUROHPC-2026-CEI-AIGF-01to establish up to seven AI Gigafactories across member states, combining up to €10 billion in EU and national public commitments designed to attract €20 billion or more in private capital. Yet bureaucracy has already forced a reset: on November 5, 2026, EuroHPC extended the deadline for medium-scale sites (Lot 1, scaling to 75,000 H100-equivalent accelerators) to February 4, 2027, kept large-scale sites (Lot 2, scaling to 100,000 accelerators) at December 3, 2026, and voided all previously uploaded draft and submitted bids due to a technical portal issue—forcing consortia to recreate and resubmit their packages.
Meanwhile, commercial champions like France's Mistral AI and a planned transatlantic merger between Germany's Aleph Alpha and Canada's Cohere (backed by Schwarz Group's STACKIT cloud) show genuine private momentum. Still, as analysts have noted, much of the underlying silicon—from GPUs inside sovereign clusters to licensed cloud stacks—remains anchored outside the EU.
When Cloud Residency Collides With Foreign Control
If Brussels wants a reality check on how deeply European enterprises lean on non-European tech, Germany's digital industry association Bitkom provides the receipts. In its Cloud Report 2026 (a representative June 2026 telephone survey of 603 German companies with 20 or more employees), 85% of respondents said Germany is too dependent on US cloud providers, up from 78% a year earlier, and 64% of cloud users said US government policies are forcing them to rethink their cloud strategy. Yet stated preference and purchasing reality remain worlds apart: 71% of surveyed companies source cloud services from US providers even though only 8% prefer them, and 43% reported that no equivalent European alternatives currently exist for their cloud requirements.
A follow-up Bitkom survey of 603 companies published in September 2026 showed the AI layer is even more lopsided. Among German firms using AI, 76% rely on OpenAI's ChatGPT, 35% use Microsoft Copilot, and 28% use Google Gemini—all three gaining share year over year. By contrast, Chinese models barely register in German enterprise surveys (2% use DeepSeek and 1% use Alibaba's Qwen), and not a single surveyed company reported using French challenger Mistral or Perplexity. As Bitkom President Ralf Wintergerst put it, German business runs almost exclusively on American AI applications.
That single-jurisdiction concentration exposes organizations to three distinct failure modes that have already played out in production over the past 16 months:
- Technical Cascading Outages: Between 23:48 PDT on October 19 and 02:40 PDT on October 20, 2025, a latent race condition between automated DNS Planner and DNS Enactor components inside Amazon Web Services' Northern Virginia (
us-east-1) region wiped out the DNS records for DynamoDB's regional endpoint (dynamodb.us-east-1.amazonaws.com). Even though other AWS regions stayed up, theus-east-1control-plane dependency rippled through EC2 instance launches, AWS Lambda, Network Load Balancers, IAM/STS authentication, and global customer apps—knocking out banking services, mobile carriers, and AI search engine Perplexity for hours until engineers manually cleared queues and throttled recovery traffic. - Sudden Export-Control Shutdowns: Just three days after Anthropic unveiled Claude Fable 5 and Claude Mythos 5 on June 9, 2026, the US Department of Commerce issued a binding export-control directive at 5:21 p.m. ET on June 12 ordering Anthropic to suspend access for all foreign nationals—both outside and inside the United States, including Anthropic's own non-citizen staff. Triggered by a reported safeguard-bypass technique flagged by Amazon researchers, Anthropic could not verify user nationalities in real time and abruptly disabled both models for every customer worldwide. The first-of-its-kind 18-day commercial AI takedown broke any enterprise app that had hardcoded Fable 5 model IDs and prompted direct Brussels–Washington talks before controls were lifted on June 30 (restoring Fable 5 globally on July 1 after a new safety classifier passed federal validation, while Mythos 5 stayed confined to vetted Project Glasswing partners).
- Extraterritorial Legal Compulsion: On June 10, 2025, Anton Carniaux, Microsoft France's director of public and legal affairs, testified under oath before a French Senate inquiry into public procurement. Asked whether he could guarantee that French citizens' data handled under public contracts would never be transmitted to US authorities without French consent, Carniaux replied: "No, I cannot guarantee it," while noting such a demand had never occurred. Because US providers fall under the 2018 US CLOUD Act, American authorities can compel disclosure of data within a US firm's possession, custody, or control regardless of which European data center stores the bits—putting providers squarely in conflict with Article 48 of the GDPR, which states that a foreign court order alone is not a lawful basis for cross-border transfer.
Cybersecurity at Model Speed: From Firefox 150 to PHP2Shell
Nowhere does dependence on frontier models bite harder than in cybersecurity, where offensive and defensive teams are now racing each other with the same class of autonomous tools.
On the offensive side, Anthropic disclosed in November 2025 that it had disrupted a mid-September 2025 espionage campaign it attributed with high confidence to a Chinese state-sponsored group that weaponized its coding assistant, Claude Code, into a quasi-autonomous agent to conduct reconnaissance, write exploits, harvest credentials, and move laterally across roughly 30 targeted corporations, financial firms, and government agencies. On the defensive side, Anthropic's Project Glasswing gave roughly 40 critical-software maintainers and major tech firms early access to Claude Mythos Preview in April 2026. Mozilla's security team pointed that preview pipeline at Firefox; on April 21, 2026, Firefox CTO Bobby Holley announced that the Firefox 150 cycle shipped fixes for 271 vulnerabilities surfaced during the evaluation, prompting him to declare that "defenders finally have a chance to win, decisively". (In a candid May 7 engineering follow-up on Mozilla Hacks, Firefox engineers clarified that those 271 AI-surfaced bugs were grouped alongside regular fuzzing and manual findings across three omnibus memory-safety CVEs—CVE-2026-6784, CVE-2026-6785, and CVE-2026-6786—plus three earlier standalone Anthropic-credited CVEs, underscoring how raw AI bug reports go through heavy triage and deduplication.)
If Firefox showed AI helping defenders clear decades of latent debt, PHP2Shell in October 2026 showed how razor-thin the margin has become when critical runtime flaws collide with AI-speed discovery. Two vulnerabilities sleeping in the PHP core—CVE-2026-109618 (memory corruption during out-of-memory cleanup) and CVE-2026-109619 (stale internal global request state across worker processes), both rated 9.2 on CVSS v4.0—enable unauthenticated remote code execution regardless of how securely an individual WordPress, Drupal, or custom app was written. Researchers from Calif (using GPT-5.6 Sol under OpenAI's Patch the Planet initiative), OpenAI, and Google-owned Wiz (leveraging multiple AI models including GPT-5.6 Sol and Gemini 4 Argon) independently uncovered and reported the flaws between August and September 2026. Wiz noted a single researcher accomplished in one week what would normally take an expert team months.
Then the embargo fractured. Amazon Threat Intelligence later reported spotting a probe against its MadPot honeypot from an anonymized IP on October 17, four days before disclosure. On October 20, CVE identifiers and a demonstration video leaked prematurely, forcing the PHP security team to ship emergency updates—PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35—a day early at 13:40 UTC on October 21, 2026. Exploit attempts surged rapidly as the European Commission, ENISA, CERT-EU, Europol EC3, and the CSIRTs Network issued an urgent joint advisory urging organizations to patch internet-facing servers immediately, isolate exposed services that could not be updated, and hunt for hidden PHP runtimes inside third-party commercial appliances and containers. Security teams should remember two vital operational truths from PHP2Shell: updating your CMS or tweaking a web application firewall (WAF) does not fix a vulnerable PHP runtime, and after patching PHP, every PHP-FPM/worker process must be restarted and containers rebuilt. As the UK's National Cyber Security Centre (NCSC) warned in its May 2025 threat assessment through 2027, the window between disclosure and exploitation is shrinking toward zero—and organizations locked out of frontier defensive AI tools risk falling permanently behind.
Persuasion, Sycophancy, and the Ballot-Box Question
Beyond uptime and zero-days, Capria highlights a quieter vector: conversational AI shaping public opinion and civic behavior. Here, peer-reviewed research published over the past year paints a nuanced—and cautionary—picture that demands careful reading without overstating what lab experiments prove.
On December 4, 2025, twin landmark studies appeared in Nature and Science. In Nature, Hause Lin, David Rand, and colleagues tested brief human–AI dialogues during the 2024 US presidential election and 2025 national elections in Canada and Poland. Chatbots instructed to advocate for a candidate shifted participants' expressed preference scores by roughly 10 points on a 0–100 scale in Canada and Poland, with a smaller shift in the US that still exceeded traditional campaign ads (roughly a third persisting a month later). Importantly, participants knew they were conversing with a partisan-prompted AI—meaning those 10-point attitude shifts do not prove 10% of real-world voters changed their ballots.
In Science, Kobi Hackenburg and researchers from the Oxford Internet Institute and the UK AI Security Institute ran three experiments involving 76,977 responses from 42,357 people, 19 language models, and 707 political issues, fact-checking 466,769 individual LLM claims. They found that persuasive power came overwhelmingly from post-training (+51%) and prompting strategies (+27%) rather than raw model size or user personalization. Models convinced people by packing responses with verifiable-sounding factual claims—yet the very tuning levers that maximized persuasiveness also degraded factual accuracy. And in a follow-up PNAS study published in October 2026 (17,950 responses from 14,779 participants), Hackenburg, Luke Hewitt, and colleagues found that persuasive AI conversations boosted real-world petition signing by up to 19.7 percentage points (Study 2; +12.8 points in Study 1)—remarkably, with no correlation between opinion shift and behavioral action, showing that information-heavy arguments move attitudes while other conversational cues drive action.
Pair those persuasion dynamics with sycophancy—the tendency of reinforcement-tuned models to agree with users even when users are wrong, famously illustrated when OpenAI had to roll back a noticeably sycophantic GPT-4o update in late April 2025—and the information ecosystem looks shaky. According to the Reuters Institute's 2026 Digital News Report, weekly use of AI chatbots for news climbed from 7% to 10% across 48 markets in a single year, reaching 17% among 18-to-24-year-olds. Yet an international study coordinated by the European Broadcasting Union (EBU) and the BBC across 22 public broadcasters in 18 countries (evaluating over 2,700 core responses from ChatGPT, Microsoft Copilot, Google Gemini, and Perplexity) found 45% of AI news answers contained at least one significant error, 31% suffered serious sourcing flaws, and 20% introduced major factual inaccuracies. A model does not need to fabricate bare-faced lies to skew public debate; subtly curating sources and validating user biases at scale—governed by training decisions made inside a handful of private labs outside Europe—carries real civic weight.
The AGI Wildcard—and an IT Resilience Playbook for Now
Every dependency headache above multiplies if artificial general intelligence (AGI) arrives—even though researchers still cannot agree on how to recognize it. A 2025 framework co-authored by Dan Hendrycks, Yoshua Bengio, and 31 collaborators scores general competence across ten cognitive domains against a well-educated adult, giving GPT-4 27% and GPT-5 57% (hampered heavily by poor continual learning across sessions without external tools). OpenAI's Charter instead defines AGI around autonomous economic work, echoing METR benchmarks showing the task duration AI agents can complete independently doubled roughly every seven months between 2019 and 2025. Meanwhile, a February 2026 Nature commentary by Eddy Keming Chen and UC San Diego colleagues argues that some frontier models already meet reasonable human-level benchmarks of general flexibility, mistakes and all. Whatever label wins out, if the first true AGI systems are built inside US firms subject to federal export controls and CLOUD Act jurisdiction, foreign governments and enterprises relying on them will hold very few cards.
To avoid waiting passively for sovereign gigafactories that will take years to pour concrete, Capria proposes a three-part "Plan B" anchored in existing EU law. It is worth separating what Brussels already requires from what remains an author recommendation:
| Resilience Pillar | Existing Legal Baseline in the EU | Author's Proposed Expansion (Not Yet Law) |
|---|---|---|
| Operational Continuity | DORA (Art. 28) mandates tested exit strategies in finance; NIS2 requires business continuity; Commission CER Guidelines (C/2026/3712, July 2026) advise manual overrides and human-in-the-loop fallbacks for critical operational technology. | Mandatory, model-specific failover drills in healthcare, energy, and public administration—measuring fraud-detection or triage speed/accuracy when a primary AI model is switched off. |
| Cyber Defense | Cyber Solidarity Act (Reg. 2025/38) establishes ENISA's EU Cybersecurity Reserve and cross-border SOCs that may use AI tools. | A shared, regularly updated European cyber-defense model trained via EUROPA/gigafactories and stress-tested in joint national exercises. |
| Democratic Transparency | AI Act requires self-assessments for systemic-risk GPAI models (backed by a Scientific Panel); DSA mandates annual public audits of very large platforms/search engines. | Standardized, regular, independent public audits of top consumer chatbots measuring factual accuracy, sycophancy, and political persuasiveness. |
For Windows IT administrators, DevOps leads, and security architects reading this from London to Berlin to Seattle, the takeaway does not depend on waiting for EU trilogues to wrap up in 2028. Treat every frontier AI endpoint like a single-region cloud database or a third-party payment gateway:
- Kill hardcoded model IDs: Put model selection behind an abstraction layer with automated health checks and pre-tested fallback routing (such as automatically dropping to an alternate provider or a self-hosted open-weight model) so an export-control order like the June 2026 Fable 5 blackout never breaks production code.
- Audit the full stack for hidden dependencies: As PHP2Shell proved, knowing your own source code is not enough—maintain Software Bills of Materials (SBOMs) to catch embedded runtimes inside vendor appliances, and separate data residency (EU servers) from data sovereignty (encryption keys and legal jurisdiction).
- Run "AI-off" continuity drills: Pick one critical workflow—SOC alert triage, fraud screening, or ticket routing—and test a manual or degraded fallback quarterly to verify your team can actually keep the lights on if an API goes dark tomorrow.
Europe does not have to beat Silicon Valley or Shenzhen to every frontier benchmark to protect its infrastructure and civic life. It—and every organization hooked on cloud-hosted intelligence—simply has to make sure no single outage, subpoena, or policy directive abroad can flip the off switch at home.
Correction (November 21, 2026): This article originally said the Stanford AI Index 2026 attributed 50 notable 2025 model releases to the United States and 30 to China; the report says 59 for the United States and 35 for China. It also said Anthropic’s leading model led its nearest US rivals by 2.7% in March 2026 Arena rankings; the 2.7% lead was over the top Chinese model. And it described the Science persuasion study as involving 76,977 participants; Science’s correction clarifies the study involved 76,977 responses from 42,357 people.
References
- Europe’s Democracy Is at Risk If It Cannot Control the AI It Depends On HackerNoon · 2026-11-20T00:00:00+00:00
- Claude Fable 5 and Claude Mythos 5 \ Anthropic anthropic.com
- anthropic.com anthropic.com