AI Pulse Daily Brief | 2026-05-07
Reading time ~11 mins
Today: the EU AI Act Omnibus trilogue stalls — 2 August 2026 deadline locks. AP barometer turns red for the first time; AFM names AI-driven markets a systemic risk. Four publishers document agentic AI moving from theoretical to confirmed real-money exploitation. Three CEOs converge on governance and integration as the durable enterprise AI value layer. Uber burns its full 2026 AI engineering budget in four months.
Top signal
EU AI Act Omnibus trilogue stalls; 2 August 2026 high-risk deadline locks as the operative planning horizon. Institute
Signal: The trilogue collapsed after twelve hours on 28-29 April over Annex I sectoral exemptions for medical devices, machinery, and connected vehicles. The follow-up trilogue is set for 13 May; even with agreement that day, Council and Parliament endorsement plus Official Journal publication cannot complete before the Cypriot Presidency hands over on 30 June, making any published relief before 2 August 2026 effectively impossible. High-risk obligations covering credit scoring, fraud detection, AML profiling, and automated customer decisions remain legally binding from that date.
Relevance: Any plan that priced in deferral to December 2027 has now accepted the risk that the next trilogue also fails — roughly thirteen weeks to land high-risk system documentation, conformity assessment, and post-market monitoring at production grade.
Consider: Freeze 2 August 2026 as the operative deadline, kill any internal slide deck still showing the December 2027 fallback, and instal a binary go/no-go gate immediately after the 13 May trilogue for any high-risk AI use case still in flight.
Security
Four independent publishers document the same week: agentic AI exploitation has moved from theoretical to confirmed real-money and real-data incidents. Institute
Signal: OECD.AI registered the first documented real-money agentic prompt-injection exploit on 4 May — a Morse-coded post on X drained roughly 175,000 dollars from a Grok-linked Bankr wallet because the AI agent decoded the adversarial instruction before applying authorisation logic. Barracuda mapped the ClawHavoc campaign that poisoned 12% of the OpenClaw skill marketplace and reached 135,000 publicly accessible agent instances in 82 countries. TechCrunch reported the LiteLLM compromise — used by roughly 36% of cloud AI environments — exfiltrated four terabytes from AI-data vendor Mercor, triggering four US class-action lawsuits. OWASP's Q1 2026 round-up codifies a new "Agentic Authorization Failure" category drawn from these incidents.
Relevance: The bank's AI vendor estate runs on a thin layer of open-source orchestration code with limited operational-risk visibility, and any agent the bank pilots in 2026 will sit on the same protocol surface that just produced four named incidents in five weeks.
Consider: Commission a four-week third-party scan for LiteLLM and downstream AI-data dependencies in the AI vendor estate, document the residual position under DORA Article 28 before the next supervisory cycle, and adopt the OWASP Q1 taxonomy as the canonical exploit classification in the June model risk review.
OECD.AI | Barracuda | TechCrunch | OWASP GenAI
Regulatory
AFM 2025 Annual Report names AI-driven capital markets as a systemic risk and demands crystal-clear accountability before widespread deployment. Authority
Signal: AFM Chair Laura van Geest classified AI-driven markets as a systemic risk in the 2025 Annual Report, with a dedicated AI capital markets supplement. The primary near-term concern is self-learning trading systems creating emergent coordination patterns that resemble collusion without intent, plus models exploiting regulatory loopholes or concealing problematic conduct. Board member Van Beusekom on the record: "Responsibilities and accountability must be crystal clear" regardless of model complexity.
Relevance: AFM is signalling the supervisory frame it will use for any AI in trading, market-making, or proprietary systems — named-individual responsibility chains and decision-logic transparency, not algorithmic explainability theatre.
Consider: Produce an AFM-aligned accountability dossier for AI in capital-markets workflows within eight weeks, naming intervention authority and human-decision points at every step from order generation to settlement.
AP turns AI Impact Barometer red for the first time; DNB and AFM confirmed as sector-specific high-risk AI supervisors for Dutch finance. Authority
Signal: The Autoriteit Persoonsgegevens published its AI Impact Barometer in red on 5 March, with four of nine indicators at critical level (up from two in the prior edition): uncontrolled deepfake spread, AI-driven fraud, chatbot psychological harm, and AI-security lag. The same publication confirms DNB and AFM as Dutch financial-sector AI supervisors and calls on government to accelerate AI Act implementation and secure structural supervisory funding.
Relevance: AP enforcement timelines historically mirror the GDPR ramp — six to twelve months from a public alarm to named-firm investigations. The bank now faces a triangulated supervisory frame: DNB on prudential, AFM on conduct and markets, AP on personal-data and cross-domain coordination.
Consider: Produce an internal exposure map of bank AI use cases to each of AP's four critical indicators before end of Q2, and prioritise the chatbot and fraud-detection portfolios — those are where AP is most likely to land first.
Perspectives
Three CEOs in seventy-two hours converge on the same thesis: foundation models commoditise, governance and integration capture durable enterprise AI value, FOMO-driven capex without ROI evidence is the structural risk. CxO voice
Signal: ServiceNow CEO Bill McDermott at Knowledge 2026 cited a real incident where an ungoverned agent deleted an entire production database in nine seconds — "governance isn't a feature, it's the whole ball game." IBM CEO Arvind Krishna at IBM Think 2026 called foundation models commodities and pointed to integration platforms and openness as the durable value layer. Goldman Sachs head of equity research Jim Covello two-year update: 95% of organisations get zero ROI on AI pilots (MIT Labs), 99% of companies reported financial losses from AI risk averaging 4.4 million dollars (EY), and "FOMO has proven a stronger incentive than poor stock performance" for hyperscaler capex.
Relevance: Three independent voices — an AI-platform CEO, a hardware-and-services CEO, and a sell-side equity research head — land on the same diagnosis. The board-level question shifts from "are we deploying enough AI?" to "where is our quantified ROI evidence and our governance kill-switch?"
Consider: Build a quantified ROI evidence pack from the bank's own production AI use cases for the September Managing Board AI review, and decide whether to formalise multi-model portability as an architecture principle before any single-vendor commitment locks in.
Fortune — McDermott | SiliconANGLE — Krishna | Fortune — Covello
Uber exhausts its full 2026 AI engineering budget in four months; GitHub Copilot ends flat-rate billing on 1 June 2026. Skeptic
Signal: Uber CTO Praveen Neppalli Naga publicly: "back to the drawing board" on AI budget planning after Claude Code adoption jumped from 32% to 84% of engineers and monthly API costs hit 500 to 2,000 dollars per engineer across a 5,000-engineer organisation. About 70% of committed code at Uber is now AI-generated; 11% of live backend updates ship without a human in the loop. GitHub moves all Copilot plans to usage-based billing from 1 June 2026; Anthropic removed Claude Code from the 20-dollar Pro tier in April.
Relevance: The flat-rate AI coding subscription era is over. Cost exposure scales non-linearly with adoption, and the bank's pilot-stage tooling assumptions will not survive contact with broad rollout.
Consider: Renegotiate or re-evaluate any flat-rate AI coding agreement before 1 June, lock multi-year cost projections to consumption-based modelling using the Uber 500-2,000 dollars per engineer per month baseline, and pair any wider rollout with a model-risk policy for AI-generated code that exceeds 10% of production commits.
Netherlands & Sovereignty
Dutch Cabinet opens consultation on the AI Act Implementation Act with eight named supervisors; consultation closes 1 June 2026. Authority
Signal: The draft Uitvoeringswet AI-verordening was published on 20 April for public consultation closing 1 June. Eight supervisors are named (AP, RDI, ILT, IGJ, NVWA, NLA, AFM, DNB). AFM and DNB are confirmed as joint supervisors for high-risk AI, prohibited practices, and transparency in the financial sector. AP is the default supervisor for Annex III systems outside sectoral authority and the cross-supervisor coordinator. RDI is the central contact point. Mandated cooperation protocols follow.
Relevance: This is the supervisory boundary the bank will operate under from 2 August 2026 onward, and the consultation window is the only time it remains negotiable.
Consider: Decide before 1 June whether to file a consultation response — particularly on the joint AFM/DNB high-risk mandate, the AP coordination role, and the AP-as-default for grey-zone use cases that touch HR, customer biometrics, or marketing analytics.
De Nederlandsche Bank signs sovereign-cloud contract with Schwarz Digits / STACKIT. Media
Signal: DNB will move major infrastructure to STACKIT (Schwarz Digits, the Lidl group's cloud arm), with Director Steven Maijoor framing it as example-setting for the Dutch financial sector. Maijoor acknowledged European alternatives are "not yet as robust or high-quality as US offerings." Announcement timed to Hannover Messe on 21 April. DNB and AFM jointly warned in October 2025 about systemic concentration on non-EU IT providers under the US CLOUD Act.
Relevance: When the prudential regulator publicly migrates its own infrastructure off US hyperscalers, that supervisory dialogue lands differently. The bank's cloud-sovereignty position is now a question DNB will ask in its next supervisory cycle, not a topic the bank can decide to defer.
Consider: Prepare a defensible cloud-sovereignty position document — naming current US-hyperscaler exposure, any sovereign-cloud migration roadmap, and the residual CLOUD Act risk — before the next DNB dialogue.
Industry & competition
ING cuts roughly 1,250 AML roles globally as AI handles 70 to 80% of standard due-diligence checks; Dutch peer cohort follows the same pattern. Media
Signal: ING is cutting roughly 1,250 jobs worldwide, primarily in its AML division (about 6,000 staff), targeting 350 million euros in cost savings by end of 2026. ING claims AI now handles 70 to 80% of standard due-diligence data points without customer involvement, compressing days into seconds. ABN AMRO is targeting 35% AML staff replacement via AI; ASN Bank is cutting 900 jobs, Triodos 250-plus.
Relevance: The Dutch supervisory and competitive baseline is being reset in public. By year-end 2026 the bank's AML cost trajectory will be benchmarked by DNB against this peer cohort, and any compliance cost story that ignores the AI-automation curve will look thin.
Consider: Benchmark current AML automation pace against the Dutch peer cohort and stress-test whether the bank's compliance cost trajectory is defensible to DNB before the next supervisory cycle.
Anthropic releases ten production financial-services agents and a Goldman-Blackstone-Hellman&Friedman delivery joint venture. Media
Signal: Anthropic launched ten specialised AI agents covering pitchbooks, account closure, and credit memo drafting, plus a new financial-analysis model for equity research and wealth management with Microsoft 365 integration (Excel, PowerPoint, Outlook). The joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs scales enterprise delivery. Existing deployments named: Goldman Sachs, Citadel, Citi, AIG. Data partnerships: Dun & Bradstreet, Moody's.
Relevance: Purpose-built financial-services tooling is moving from pilot to production faster than internal builds can match. The bank's automation roadmap will start being measured against externally available agents that already cover pitchbook, KYC, and credit-memo workflows out of the box.
Consider: Evaluate the ten Anthropic agents against existing internal automation roadmaps before end of Q2, and decide whether to engage the joint venture's delivery channel directly or wait for the consultancy market to reprice.
Innovation
Anthropic signs a 300 megawatt, 220,000-GPU compute deal with SpaceX from Colossus 1 in Memphis; Claude usage limits raised immediately. Vendor
Signal: SpaceX makes 300-plus megawatts and 220,000 NVIDIA GPUs from Colossus 1 available within the month — capacity originally built for xAI's Grok models, post the SpaceX-xAI merger earlier this year. Anthropic raised Pro and Max usage limits immediately and named a multi-gigawatt space-based compute development as joint interest with SpaceX.
Relevance: The headline capacity-ceiling risk for any bank standardising on Claude in the next 12 to 18 months has materially eased. This is a positive supply-side signal for vendor strategy, not a regulatory or competitive one.
Consider: Treat this as relief on the Claude-availability tail risk in vendor-strategy committee, but do not relax multi-model portability discipline — the supply-side signal is favourable, the geopolitical and concentration risks unchanged.
Research
BCG, IBM Institute for Business Value, and McKinsey publish in the same week converging on the same thesis: CEO-board AI alignment is structurally weak and organisational readiness is the binding constraint. Advisory
Signal: BCG (625 CEOs and boards) finds 61% of CEOs say their boards are rushing AI — inverting the usual risk narrative — with AI spend doubling to 1.7% of revenues in 2026 and 30%-plus of that committed to agentic AI. IBM IBV (2,000 CEOs across 33 geographies) reports Chief AI Officer prevalence jumped from 26% to 76% in one year and organisations expect 48% of operational decisions to be AI-made without human involvement by 2030. McKinsey (10,000-plus senior executives) finds 72% say their organisations are not fully ready and 75% fail to build cultures capable of navigating the transition.
Relevance: Three independent data sets converge on the same finding: governance and organisational readiness, not technology, is what separates AI value-capture from AI capex burn. The bank's Managing Board and Supervisory Board face a structural alignment question that the data now makes explicit.
Consider: Schedule a joint Managing Board / Supervisory Board AI risk-appetite and pacing dialogue, and explicitly decide and document the bank's AI accountability structure — CAIO mandate scope relative to CRO, CTO, CDO — before the year-end strategic review.
On the radar
- EU AI Act Article 50 transparency obligations enter operative force on 2 August 2026, requiring machine-readable marking of AI-generated audio, image, video, and text outputs; Code of Practice final draft expected June 2026 (penalties up to 15 million euros or 3% of global turnover). European Commission
- HSBC appoints David Rice as inaugural Chief AI Officer effective 1 April, separate from CTO at C-suite level — peer cohort normalises the CAIO design. HSBC
- Stanford 2026 AI Index reports agent task-success jumped from 20% to 77.3% in one year and cybersecurity agents from 15% to 93%; 47 countries have AI legislation but only 12 have enforcement mechanisms. Stanford HAI
- ASML Q1 2026 China share of system sales dropped from 36% (Q4 2025) to 19% as export controls take full effect; full-year guidance still raised to 36-40 billion euros. ASML
- PYMNTS: 63% of firms have identified Know-Your-Agent risks in loan applications and half have already incurred losses from adversarial bots; identity stacks designed around human counterparties are breaking. PYMNTS
- CBS: Dutch micro-enterprise AI adoption doubled in two years to 13.8% of firms generating 17.9% of sectoral turnover — SME bifurcation visible in turnover data. CBS