Other
AI: capabilities, regulation, labour
Updated 15m ago·Update
Models are learning faster than the institutions behind them can adapt. The thread tracks frontier-model capability jumps, the AI Act and its enforcement, labour-market impact, and infrastructure (chips, energy, water).
Latest update
First AI Act Systemic-Risk Probe Launched
The European Commission formally launches its first systemic-risk investigation under the AI Act, targeting a frontier model with advanced autonomous cyber capabilities. The probe will test the regime's ability to compel developers to provide post-deployment evaluations and red-teaming data.
State of play
The AI Act's enforcement engine is now live. The European Commission has opened its first systemic-risk investigations into two frontier model developers, a decisive move from rule-writing to active oversight. This action is backed by a parallel institutional build-out, with Germany and France standing up national safety institutes to conduct adversarial testing. Their findings will feed the EU's push to anchor regulation in dynamic capability benchmarks, a pivot solidified by a new draft rule swapping the static compute trigger for a test based on autonomous cyber and coding prowess. The immediate precedent rests on the outcome of the first probes: whether the Commission can successfully compel the disclosure of post-deployment data to validate the external benchmarks that now define the regulatory frontier.
Chronicle
View historyItaly and Spain create joint taskforce on AI labour impact
Italy and Spain form a joint taskforce on AI and employment to coordinate responses to automation-driven restructuring. The taskforce will collect data on job impacts and explore EU-level funding for reskilling, as unions push for consultation clauses in collective bargaining.
First systemic-risk probe targets two frontier models
The European Commission launches its first formal investigation under the AI Act's systemic-risk provisions, targeting two unnamed frontier models. The probe, coordinated by the EU AI Office, will assess compliance with risk management, cybersecurity, and incident reporting obligations.
Capability-based triggers fast-tracked into AI Act rules
In parallel, the Commission begins revising the AI Act's delegated acts to embed capability-based triggers, such as autonomous cyber horizons, as the primary criteria for defining systemic-risk models. This move aims to make regulatory thresholds explicitly adjustable based on model evaluation data.
Germany launches federal AI safety lab
Germany establishes a federal AI Safety and Security Lab under its Federal Office for Information Security (BSI) to test high-risk and systemic-risk models. The lab will collaborate with the EU AI Office, focusing on models used in industrial control, energy, and healthcare.
France opens national AI evaluation centre
France inaugurates a National AI Evaluation Centre tasked with standardized safety testing for both large foundation models and smaller systems embedded in public-sector use cases. It will support regulators and participate in joint testing with international partners.
Nordic agreements pivot AI rollout to redeployment
New sectoral agreements in Sweden, Denmark, and Finland commit large employers to prioritize retraining and role changes over layoffs when introducing advanced AI. The deals include joint monitoring committees to review productivity and employment effects.
UK Safety Institute confirms acceleration in autonomous cyber capabilities
The UK AI Safety Institute publishes updated longitudinal data, confirming that the doubling time for the complexity of autonomous cyber tasks frontier models can reliably complete remains at or below 4.7 months. The institute's analysis of models released in early 2026 shows the trend holding as context windows and tool-use improve.
First AI Act Systemic-Risk Probes Begin
The European Commission launches the first formal systemic-risk investigations under the AI Act into two unnamed frontier model developers. The probes test the Act's new powers to demand post-deployment logs, incident reports, and red-team data to verify compliance with continuous monitoring and emergent-risk reporting obligations.
Germany, France Launch National Safety Institutes
Germany and France establish national AI safety institutes with mandates to adversarially test frontier models. Germany's Bundesinstitut für KI‑Sicherheit will focus on models in critical infrastructure, while France expands its Paris AI lab into a national evaluation hub, both aiming to feed results into the EU's systemic-risk classification process.
Autonomous Cyber Capability Growth Confirmed
New data from the UK AI Safety Institute confirms the complexity of autonomous cyber tasks frontier models can complete continues to double roughly every five months. Independent benchmarks show these models can now compress year-long manual penetration tests into weeks, executing multi-step intrusion chains autonomously.
EU Proposes Dynamic Risk Test to Replace FLOP Threshold
The Commission's AI Office circulates a draft implementing act proposing to replace the AI Act's static compute threshold with a dynamic risk test based on real-world capability benchmarks, such as autonomous cyber offence. The shift is explicitly driven by the empirical acceleration timelines published by safety institutes.
Brytyjski regulator nakazuje Google umożliwienie wydawcom rezygnacji z AI w wynikach wyszukiwania — światowy precedens
Brytyjski Urząd ds. Konkurencji i Rynków nałożył na Google nowe wymogi dotyczące postępowania, zmuszając giganta technologicznego do umożliwienia wydawcom blokowania ich treści przed funkcjami wyszukiwania opartymi na sztucznej inteligencji, bez uszczerbku dla ich pozycji w standardowych wynikach wyszukiwania.
UK safety institute reports renewed acceleration in autonomous cyber capabilities
The UK AI Safety Institute releases new data indicating the effective time horizon for models to complete autonomous cyber tasks has shortened, with capabilities on complex software tasks doubling in roughly four months. The findings show models can sustain multi-step intrusion sequences with limited oversight.
EU Commission signals readiness to adjust AI Act thresholds
Senior European Commission officials state they are prepared to revise the AI Act's frontier-model thresholds and systemic-risk designations. The AI Office is studying whether current compute thresholds and testing templates are sufficient to capture powerful autonomous-agent behaviours, signalling a move toward more capability-based criteria.
Commission moves to tighten frontier-model thresholds
The European Commission begins drafting a delegated act to formally update the AI Act's definition of 'systemic risk' frontier models. The planned change would shift the framework toward more dynamic, capability-focused criteria, allowing regulators to impose obligations on models with advanced autonomous capabilities, even if their training compute falls below the original 10^25 FLOP threshold.
France and Germany establish national safety institutes
Germany advances legislation to create a federal AI Safety and Security Institute, while France expands its national AI hub into a dedicated frontier model evaluation centre. Both institutes are mandated to conduct independent safety testing and red-teaming, feeding technical evidence into the EU AI Office's systemic-risk assessments.
Transatlantic talks begin on AI chip export controls
US and EU policymakers begin discussions on coordinating tighter export controls for advanced AI chips, aiming to restrict access to high-end hardware for training frontier models in countries deemed high-risk for military or cyber misuse.
EU opens first systemic-risk investigation under AI Act
The European Commission initiates its first formal probe into a frontier AI model designated as posing a potential systemic risk. The investigation focuses on whether the model's autonomous cyber and code-generation capabilities breach obligations for risk monitoring and mitigation, testing the AI Office's powers to demand post-deployment evaluations and incident reports.
Technical data quantifies rapid acceleration in autonomous cyber capabilities
Analysis from the UK AI Safety Institute estimates the length of autonomous cyber tasks frontier models can complete has been doubling roughly every 4 to 5 months since late 2024. This data is cited by EU officials as concrete evidence that fixed compute thresholds are obsolete, bolstering the case for dynamic, capability-based regulatory triggers.
Security research underscores near-term autonomous threat
Reports from security researchers and the Frontier Model Forum indicate current models can autonomously chain exploits and adapt to defenses, completing vulnerability discovery work comparable to a year of manual penetration testing in under three weeks. This reinforces policy arguments for capability-based risk designations.
Regulators advance capability-based definitions for frontier AI
Policy memos describe a clear shift in the EU and some US states toward defining systemic-risk AI by demonstrated capabilities like advanced cyber functions, moving beyond static training-compute thresholds. The EU is preparing secondary legislation to operationalize this flexible approach within the AI Act.
UK competition regulator sets precedent on AI and publisher content
The UK's Competition and Markets Authority orders Google to allow publishers to opt their content out of AI-powered search features without penalty to their ranking in standard search results. This creates a global precedent for separating AI training and indexing from core platform services.
EU AI Office launches first systemic-risk investigation
The European Commission's AI Office has initiated its first formal probe under the AI Act, examining a frontier-model developer's pre and post-release safety testing, capability evaluations, and incident reporting. National AI safety institutes are providing technical support, creating a test-bed for future EU enforcement workflows.
France inaugurates national AI safety institute
The French government has officially launched a national AI safety institute with a multi-year budget. Its mandate includes technical evaluation of frontier models, focusing on capabilities like autonomous cyber operations, to feed findings into the EU AI Office's systemic-risk assessments.
Germany scales federal AI safety centre for EU audits
Germany has expanded its federal AI safety and trust centre, giving it a formal mandate to conduct technical audits of systemic-risk models designated under the AI Act. The centre is recruiting specialists in robustness and cyber security to support EU-level capability evaluations.
Commission drafts EU-wide controls on AI chip exports
The European Commission is preparing a proposal for coordinated controls on exports of advanced AI chips and large training clusters, aiming to align with US and Japanese restrictions. The draft framework also explores licensing for large cloud-based compute allocations to non-EU entities.
AI Office issues guidance linking capabilities to risk designation
The EU AI Office has published guidance clarifying that systemic-risk designation can be based on demonstrated capabilities (like autonomous cyber operations) beyond just training compute. It outlines expectations for developer-supplied evaluation results, tying testing practices directly to compliance.
UK institute reports sustained rapid gains in AI cyber autonomy
The UK AI Safety Institute reports that the length and complexity of cyber tasks frontier models can complete autonomously continues to grow, with a doubling time for reliable tasks near 4–5 months. European regulators cite this as evidence for dynamic, capability-based oversight.
Microsoft Build 2026: chip kwantowy z celem na 2029, zapowiedź noszonych urządzeń AI i pogłębienie partnerstwa z Nvidią
Podczas konferencji deweloperskiej Build 2026 Microsoft zaprezentował nowy chip kwantowy z celem na 2029, zapowiedział koncepcje noszonych urządzeń zasilanych AI oraz pogłębił partnerstwo sprzętowe z Nvidią.
First systemic-risk probe launched under AI Act
The EU AI Office opens its first formal investigation into a frontier foundation model, using capability-based criteria like autonomous cyber performance to assess systemic risk. National authorities in three member states are called to assist with technical testing.
National AI safety institutes expand for EU Act
France and Germany announce expansions of their national AI safety institutes, explicitly aligning their mandates with the EU AI Act's needs for red-teaming, safety testing, and post-market monitoring of general-purpose models.
AI data centre boom strains EU infrastructure
Energy and water regulators across the EU warn that the projected boom in AI data centres is straining local power grids and water systems, especially in southern Europe. The Commission examines whether to update sustainability reporting rules for large data centres.
EU moves to tighten AI chip export controls
The European Commission proposes updated dual-use export controls, coordinated with the US and Japan, to require licenses for advanced AI accelerators and high-bandwidth memory destined for jurisdictions of concern. Negotiations begin over potential exemptions for domestic semiconductor firms.
Autonomous cyber capability growth accelerates
New analysis from the UK AI Safety Institute finds the time horizon for frontier models to perform complex, autonomous cyber operations is shortening faster than expected, with capabilities now doubling every 4.7 months. This acceleration outpaces previous projections.
AI automation cited in new wave of job cuts
Major employers in Germany, the Netherlands, and Spain announce white-collar job cuts linked to productivity gains from generative AI tools automating reporting, legal drafting, and customer support. Trade unions call for faster EU guidance on labour law in AI-driven reorganisations.
Copyright battles over AI training data intensify
Coordinated copyright lawsuits are filed or expanded in France, Germany, and Italy against OpenAI and Anthropic, alleging unlicensed use of protected text and images for training. The suits test the interaction of EU copyright law with the AI Act's new transparency duties.
Trump podpisuje okrojoną dyrektywę w sprawie AI, dając rządowi 30-dniowy dobrowolny przegląd zaawansowanych modeli
Prezydent Trump podpisał we wtorek zarządzenie wykonawcze, w którym prosi firmy z branży AI o udostępnianie swoich najpotężniejszych modeli do dobrowolnego przeglądu rządowego na maksymalnie 30 dni przed publiczną premierą, wycofując się z wcześniejszych projektów, które nakazywałyby 90-dniowe okno.
AISI Quantifies Explosive Acceleration in Autonomous AI Cyber Capabilities
The UK AI Safety Institute (AISI) publishes analysis showing the 'task-length horizon' for frontier models to autonomously complete cyber operations is doubling every 4.7 months, a pace comparable to the most explosive software task improvements. This acceleration, tracked since late 2024, places cyber capabilities on a months-not-years advancement timeline.
EU AI Office launches first formal investigations under AI Act
The European Commission's AI Office has initiated its first formal probes into potential breaches of the AI Act by frontier-model providers. The investigations focus on obligations around training data transparency, systemic-risk assessments, and red-teaming, establishing reference cases for enforcement across the single market.
Commission issues first detailed guidance on systemic-risk model designation
The European Commission has published its first guidance on classifying AI models as posing a 'systemic risk'. It elaborates on the 10^25 FLOPs compute threshold and introduces capability-based criteria, such as autonomous cyber and code-generation skills, granting the AI Office wide discretion to open assessments.
Germany and France formally establish national AI safety institutes
Germany and France have launched their national AI safety institutes to support EU-level oversight. The German institute focuses on technical evaluations of high-risk models, while France's body will advise sectoral regulators and conduct independent capability assessments, serving as potential templates for other member states.
Benchmarks reveal frontier models' rapid gains in autonomous coding and cyber operations
Recent technical evaluations, including from the UK's AI Safety Institute, show frontier AI models achieving sharp performance jumps on complex software engineering and cyber-security tasks. Models can now autonomously chain vulnerabilities and execute multi-step attack plans, informing regulatory scrutiny on capability-based risk thresholds.
EU and US expand AI chip export controls, widening scope beyond China
The EU and United States have coordinated to tighten export controls on advanced AI chips and chipmaking tools. The expanded measures extend licensing requirements to a broader range of accelerators and restrict transfer of advanced packaging technologies, aiming to manage security risks and prevent circumvention.
AI automation cited as driver in new wave of tech and media layoffs across Europe
A new wave of job cuts in technology, media, and business-services firms explicitly cites generative AI automation as a driver. Investment is shifting from routine content production, support, and clerical roles to AI tooling and engineering, increasing displacement risk for older workers and those in non-specialist positions.
European media and authors file expanded copyright lawsuits against OpenAI and Anthropic
European media groups and authors have filed new or amended lawsuits against OpenAI and Anthropic in courts in France, Germany, and the Netherlands. The claims allege copyright violation from training on scraped content and argue that AI Act compliance should include detailed disclosure of training data sources.
UK AI Safety Institute and EU bodies deepen cooperation on frontier model testing
Collaboration between the UK's AI Safety Institute and EU institutions has intensified, involving sharing red-teaming methodologies, co-developing benchmarks, and exploring joint testing of models. This creates an emerging network of public AI safety institutes spanning the UK, Germany, France, and EU-level bodies.
Energy and water demands of AI datacentres trigger local regulatory responses in EU
Growing concern over the resource consumption of AI-focused datacentres is prompting regulatory responses in several EU member states. Authorities are tightening planning rules, introducing reporting obligations, and requiring operators to use recycled water or invest in grid upgrades, targeting the physical footprint of the AI boom.
Analysis confirms EU AI Act's capability-focused systemic risk powers are now active.
A Third Way memo clarifies that the European Commission and EU AI Office can now designate models as systemically risky on a case-by-case basis, using criteria beyond pure compute thresholds. This activates a more flexible, capability-focused enforcement tool within the Act.
UK AI Safety Institute reports frontier model cyber capability horizon doubling every 4.7 months.
The institute's latest assessment shows the time horizon for frontier models to achieve 80% reliability on cyber tasks has accelerated, now doubling every 4.7 months since late 2024, down from an earlier 8-month estimate. This indicates a quickening pace in a key autonomous capability.
UK AI Safety Institute publishes benchmark showing frontier models' autonomous cyber capability doubling every 4.7 months.
The institute's analysis, based on evaluations since late 2024, provides a concrete metric for the acceleration regulators must now race against. It underscores that evaluation and oversight timelines are shrinking dramatically.
Security firm Unit 42 reports frontier AI models can autonomously complete the equivalent of a year's manual penetration testing in less than three weeks.
This finding from Palo Alto Networks illustrates the practical, near-real-time impact of the capability acceleration quantified by the UK institute, directly relevant to cybersecurity risk assessments.
Policy analysis argues static definitions of frontier AI are obsolete, highlighting the EU AI Act's move towards dynamic, capability-based designation of systemically risky models.
The memo points to the Act's mechanism for regulators to designate models based on their capabilities as a necessary adaptation to rapid model jumps and emergent abilities.
EU AI Office opens first systemic-risk investigation into a frontier model
The EU AI Office has launched its first formal probe into a frontier AI system designated as posing potential systemic risk, testing its new oversight powers. The investigation focuses on the adequacy of the model's safeguards against autonomous cyber misuse and critical-infrastructure interference, and the pace of its capability evolution.
Germany establishes a national AI safety institute
Germany's federal government has approved the creation of a national AI safety institute, mirroring the UK model. It will be tasked with independent evaluations of frontier models, red-teaming for misuse risks, and advising on technical standards for AI Act compliance.
France expands INRIA and ANSSI mandates into a joint AI safety testbed
The French government has broadened the mandates of research institute INRIA and cybersecurity agency ANSSI to build a joint testbed for evaluating high-risk and frontier AI systems. The platform will focus on autonomous cyber operations and provide technical input for the EU AI Office.
EU Commission drafts fast-track procedure to update AI Act systemic-risk criteria
The European Commission is preparing a delegated act to allow the AI Act's systemic-risk criteria and test methodologies to be updated rapidly, responding to findings that autonomous cyber capabilities are doubling in months.
Palo Alto Networks research shows frontier models compress a year of pentesting into weeks
New research confirms that frontier AI models can autonomously identify vulnerabilities and chain exploits, performing the equivalent of a full year of manual penetration testing in under three weeks, dramatically shortening the window for defenders.
European tech and media groups expand copyright suits against AI firms
Coalitions of publishers and media groups in several EU countries have expanded legal actions against OpenAI and Anthropic, testing new AI Act provisions requiring training-data summaries and copyright policies.
UK AI Safety Institute and EU AI Office plan shared test protocols for autonomous cyber risk
The UK AISI and the EU AI Office have begun drafting shared test protocols for measuring autonomous cyber-intrusion capabilities in frontier models, aiming to harmonise benchmarks and ensure consistent interpretation of evaluation results.
EU moves toward coordinated export controls on advanced AI chips
Following US actions, EU officials have begun talks with Washington and Tokyo on aligning controls for the most powerful accelerators used to train frontier models, exploring an EU-wide regime while navigating divisions among member states.
AI-related layoffs spread across European white-collar sectors
Major firms in finance, media, and business services have announced layoffs explicitly citing efficiency gains from generative AI, affecting back-office and analytical roles, while demand for AI specialists grows.
AI Index 2026 confirms PhD-level science capability but persistent reliability gap
The Stanford AI Index 2026 report documents a roughly 30% year-on-year improvement for frontier models on the Humanity's Last Exam benchmark, with performance now meeting or exceeding human baselines in PhD-level science questions, multimodal reasoning, and competition mathematics. However, the same systems fail approximately one in three attempts on structured enterprise benchmarks and struggle with basic perception, highlighting a critical divergence between peak capability and day-to-day reliability. This reliability chasm presents a core challenge for both regulatory oversight, which must manage the risks of superhuman performance, and commercial deployment, where uneven accuracy between 60-90% across domains like tax and legal reasoning remains a barrier.
UK AI Security Institute reports frontier models' cyber task horizons doubling every 4.7 months
New analysis from the UK's AI Security Institute (AISI) finds the 'task horizon' for frontier AI models performing cyber operations at 80% reliability has been doubling approximately every 4.7 months since late 2024. This pace is faster than the institute's prior estimates and aligns closely with the rapid improvement rates seen in general software tasks, signalling an accelerated advancement in autonomous cyber capabilities.
OpenAI Hit with Fresh EU Copyright Lawsuits
OpenAI faces new copyright lawsuits in the EU from publishers and authors' societies, challenging both training-data use and the reuse of generated content.
Anthropic Faces EU Legal Challenges Over Data Scraping
Anthropic and other AI labs face European legal action over alleged scraping of copyrighted text and code from EU-hosted sites for training frontier models.
Regulators Warn AI Data-Centres Risk Grid and Water Shortages
European energy and water regulators warn that rapid AI data-centre expansion risks local electricity and water shortages, urging updated planning rules and efficiency standards.
AI-Driven Layoffs and Restructuring Hit European Tech Sector
European tech and media firms announce AI-linked layoffs and restructuring, citing efficiency gains from generative AI. Unions and labour ministries call for stronger social dialogue and reskilling funds.
Unions and Employers Negotiate AI Workplace Rules
EU social partners negotiate framework agreements on generative AI use in the workplace, seeking to define limits on surveillance, algorithmic management, and deskilling.
First EU Systemic-Risk Probe Begins
The EU AI Office launches its first formal systemic-risk investigation into a frontier model provider, initiating active enforcement under the AI Act. The case is intended to serve as a template, testing demands for transparency on training data, compute, and monitoring for models near the high-risk threshold.
Autonomous Cyber Skills Accelerate
New evaluations from the UK AI Safety Institute and partners confirm that the autonomous cyber and coding capabilities of frontier models continue to double every four to five months. This acceleration compresses the timeline for models to move from 'general-purpose' to 'systemic-risk' classification between regulatory compliance cycles.
Studies Expose Pervasive Jailbreak Vulnerabilities
Security research reveals that leading frontier models remain widely susceptible to sophisticated jailbreak and prompt-injection attacks, even with safety filters. The findings are being cited by regulators as evidence for the need for continuous, not one-off, adversarial testing under the AI Act's compliance regime.
EU Refines Dynamic Model Definition
EU policy notes clarify that the definition of frontier models for systemic-risk designation will be dynamic, based on both compute thresholds (around 10^25 FLOP) and demonstrated capabilities. This allows regulators to re-classify models as their performance jumps, a response to the faster-than-anticipated pace of advancement.
EU AI Office launches landmark probe into frontier model's systemic-risk compliance
The Office opens its first formal investigation under the AI Act's systemic-risk regime, examining a major provider's latest model for potential gaps in risk management, cybersecurity safeguards, and required transparency documentation. The case is framed as setting a precedent for using capability-based, not just compute-based, criteria to designate high-risk models.
UK AI Safety Institute quantifies exponential acceleration in autonomous cyber skills
An updated analysis confirms the 'effective time horizon' for complex AI-driven cyber tasks has been doubling approximately every 4.7 months since late 2024. This rapid gain in persistence and task length is noted to quickly obsolete existing security benchmarks and regulatory testing frameworks.
Cybersecurity researchers warn frontier models can compress year-long pentests into weeks
Palo Alto Networks' Unit 42 reports that state-of-the-art models can now autonomously identify vulnerabilities and chain exploits to achieve in less than three weeks what typically requires a year of manual penetration testing. This highlights the present technical capability for near-real-time, AI-driven hacking campaigns.
First systemic-risk probe launched by EU AI Office
The EU AI Office opens its first formal investigation into a frontier model provider, aiming to determine if its models qualify as a systemic risk under the AI Act's dynamic, capability-based powers. This live test of enforcement agility directly targets the provider's risk management and documentation.
Autonomous cyber capabilities doubling every 4-5 months
The UK AI Safety Institute updates its analysis, reporting that the complexity of cyber tasks frontier models can complete autonomously is doubling every 4-5 months. This acceleration means work previously requiring a year of manual effort can now be achieved in weeks.
Year of pentesting condensed to weeks by AI
Security researchers confirm that frontier models can autonomously perform the equivalent of a full year of manual penetration testing in under three weeks, chaining low-severity bugs into critical exploits. This underscores the dual-use risk of democratizing cyber capabilities at unprecedented speed.
EU and US Coordinate on AI Chip Export Controls
EU and US officials move towards coordinated export controls on advanced AI accelerators and large training clusters, aiming to harmonise thresholds and prevent regulatory arbitrage.
NL and Italy Tighten Scrutiny on AI Chip & Data-Centre Investments
The Netherlands and Italy tighten investment screening rules for AI chip manufacturing and large data-centre projects, citing national security and grid-stability concerns.
First Systemic-Risk Probe Targets Frontier Model's Cyber Capabilities
The European Commission's AI Office opens its first systemic-risk investigation under the AI Act, targeting Anthropic's 'Mythos' frontier model. The probe will assess whether the model's risk management, red-teaming, and access controls are adequate for a system whose autonomous cyber capabilities are advancing rapidly.
AI Office Sets 4-6 Month Review Cycles for Frontier Models
The Commission publishes its first systemic-risk guidance for frontier AI, outlining expectations for continuous monitoring, incident reporting, and safety evaluations. It aims for review cycles of 4-6 months, acknowledging that cyber-relevant capabilities can double faster than traditional regulatory timelines.
Germany Establishes National AI Safety Institute
Germany launches a Federal AI Safety Institute to independently evaluate high-risk and systemic-risk AI models used within its jurisdiction. The institute will focus on red-teaming frontier systems for cyber, critical infrastructure, and democratic-process risks, complementing EU-level oversight.
France Begins Red-Teaming Models in Critical Sectors
France activates a national AI evaluation centre to red-team foundation models deployed in critical sectors like healthcare and finance. Its findings will be shared with the EU AI Office and sectoral regulators.
UK Expands AI Safety Monitoring to Real-Time Deployment
The UK AI Safety Institute expands its mandate to include real-time monitoring of deployed frontier models, particularly in sensitive domains like cybersecurity. It plans to share methodologies with EU counterparts to support enforcement.
Creative Industries Push for AI Training Registries and Funds
Collective EU creative-industry groups campaign for mandatory AI training registries and sector-specific compensation funds for rights-holders.
Commission Considers Fast-Track Tools for AI Act Enforcement
The European Commission explores fast-track enforcement tools under the AI Act, including interim measures, to respond to capabilities advancing on sub-year timescales.
EU AI Office opens first AI Act investigations
The European Commission's AI Office launches its first formal investigations into potential breaches of the EU AI Act by several large AI providers, focusing on transparency, safety documentation, and systemic-risk obligations.
Operational pause in major developments
A news cycle passes without major, verifiable developments in frontier AI capabilities, significant regulatory enforcement actions, or substantial shifts in the EU's oversight infrastructure, indicating a period of operational consolidation.
Industry report exposes major gap in AI safety benchmarks
Cisco's AI threat intelligence unit publishes a report revealing frontier models are highly vulnerable to sophisticated, multi-turn jailbreak attacks, with success rates as high as 88%, far exceeding what standard safety benchmarks indicate.
UK safety body warns of rapidly accelerating autonomous AI cyber capabilities
The UK AI Safety Institute reports that the length of cyber tasks frontier models can autonomously complete at high reliability has been doubling every 4–5 months since late 2024, a significant acceleration from previous estimates.
Safety evaluators warn access to frontier AI is outpacing risk assessment
METR's Frontier Risk Report for early 2026 flags that expanding API access to powerful models is outpacing rigorous evaluation, increasing systemic risk as dangerous capabilities reach more users faster than safety infrastructure can adapt.
Benchmark reveals steep one-year jump in autonomous AI agent proficiency
The OSWorld benchmark shows a dramatic one-year leap in AI agent performance, with task success rates on complex computer operations jumping from ~12% to ~66%, underscoring a discontinuous improvement in autonomous action.
Germany and France establish national AI safety institutes
Germany has created a national AI safety institute with a mandate for independent testing of advanced models, mirroring similar bodies in the UK and US. France has expanded an INRIA-led hub into a broader national safety and evaluation platform. Both aim to provide technical assessments to support EU-level enforcement and national sectoral regulators, responding to evidence of shrinking timelines for dangerous capabilities.
Evaluators report sharp jumps in general computer autonomy
Independent evaluators METR and OSWorld report that frontier models' ability to autonomously operate computers—navigating GUIs and completing multi-step tasks—has improved markedly. This shift from partial assistance to near-end-to-end task completion highlights how real-world autonomy emerges when models are connected to tools, a risk factor traditional benchmarks understate.
EU and US align on tighter AI chip export controls
The US has updated export controls on high-end AI accelerators, with the EU aligning aspects of its dual-use regime. The measures target data-centre-class GPUs to slow military and surveillance applications by strategic competitors. Chipmakers warn that frequent rule changes complicate supply-chain planning in Europe, where demand for AI compute is already constrained.
EU launches first systemic-risk probe under AI Act
The European Commission's AI Office has initiated its first formal investigation into a frontier model designated as a systemic risk. The probe focuses on whether the provider's safety testing and monitoring for high-risk dual-use capabilities, like autonomous cyber operations, are sufficient. This case will set a crucial precedent for how the EU enforces its compute-based thresholds and the clause allowing reclassification based on capabilities alone.
UK data reveals months-scale gains in AI cyber autonomy
The UK AI Safety Institute has published new results indicating frontier models' autonomous cyber capabilities are roughly doubling on a timescale of months. Recent models show a dramatically improved ability to execute longer, more complex hacking tasks by chaining exploits and adapting to defences. This quantitative evidence underscores the rapid pace at which institutional responses are falling behind.
European finance and tech firms cut jobs citing AI gains
Major banks and insurers in Germany, France, Spain, and the Netherlands have announced thousands of back-office job cuts, explicitly linking them to the deployment of generative AI. Simultaneously, tech and media companies across Europe are implementing layoffs or hiring freezes in functions like marketing and coding. This marks a concrete, visible labour-market impact as AI reshapes white-collar workflows faster than reskilling programmes can respond.
This week
- EU launches first AI Act systemic-risk probes into two frontier models.
- Germany and France establish national AI safety testing institutes.
- UK data confirms 5-month doubling time for autonomous cyber tasks.
- Commission proposes dynamic risk test to replace static compute threshold.