Which system, use, population and consequence does the evidence cover?
- Evidence rule
- Separate law, standards, guidance, voluntary commitments and research.
FUURAA™ · 2026 AI Frontier Technology Radar
Safety, Evaluation, Policy & Society
Connect technical evaluation, operational controls, standards, public policy and social consequences.
Watch thesis
Responsible AI is not a statement of intent; it is a chain of scoped evidence, enforceable controls, accountable decisions and usable recourse.
Applicability boundaryThis page is research synthesis, not legal advice, certification, audit opinion or a determination of compliance.
This page separates canonical sources, FUURAA original summaries and analysis. Inclusion does not imply partnership, endorsement or approval.
What this lens tracks
Declare the observation boundary before admitting a new development into the evidence chain.
- 01safety research, evaluations and incident evidence
- 02standards, regulation and institutional governance
- 03labour, information integrity and public impact
Three evidence questions
Ask not only what happened, but how far the evidence travels.
Are controls technically enforced and tested under failure?
- Evidence rule
- Date every source and state jurisdiction and applicability.
Can affected people understand, challenge and obtain remedy?
- Evidence rule
- Do not infer compliance or safety from citation alone.
Verified primary sources
10 directly relevant 2026 source records for this lens.
Each record states its publication date, source organisation, evidence type, FUURAA original summary and known boundary.
16 July 2026 · U.S. Department of Energy / Lawrence Livermore National LaboratoryCESER Releases New Testbed to Advance LLM and Agentic AI Evaluation for Critical Infrastructure
Evidence statusPrimary source · Product release
Stormbreaker is a dynamic testbed for evaluating language models and agents in power-system and operational-technology environments before operational deployment.
Open canonical source ↗2 July 2026 · UK AI Security InstituteMore compute, more capability: Why AI agent evaluations need to account for test-time computeEvidence statusPrimary source · Research
The findings indicate that fixed compute budgets can understate agent capability and argue for reporting capability as a curve over test-time compute rather than a single score.
Open canonical source ↗10 June 2026 · European Commission AI OfficeCommission publishes Code of Practice on marking and labelling AI-generated contentEvidence statusPrimary source · Standard
The voluntary code sets practical marking and disclosure measures ahead of EU AI Act transparency obligations applying from 2 August 2026.
Open canonical source ↗8 June 2026 · UK AI Security InstituteRealityTest: Do AI systems disclose their identity when asked?Evidence statusPrimary source · Research
RealityTest uses multilingual, human-authored identity probes to assess whether text and speech systems disclose that they are AI, finding strong sensitivity to phrasing, context and system instructions.
Open canonical source ↗20 May 2026 · Singapore IMDAUpdated Model AI Governance Framework for Agentic AIEvidence statusPrimary source · Policy
Singapore's updated framework adds deployment cases and practices for multi-agent systems, third-party agents, automation bias, bounded autonomy and meaningful human accountability.
Open canonical source ↗15 May 2026 · Council of EuropeEuropean Union ratifies the Council of Europe Framework Convention on Artificial IntelligenceEvidence statusPrimary source · Policy
EU ratification advances the first legally binding international AI treaty centred on human rights, democracy and the rule of law.
Open canonical source ↗20 April 2026 · Singapore IMDA / Enterprise SingaporeSingapore Champions New Global AI Testing Standardisation EffortsEvidence statusPrimary source · Standard
Singapore proposed ISO/IEC 42119-8 to standardise testing methodology for generative-AI systems.
Evidence boundaryProposed international standard; not a final published ISO standard.
Open canonical source ↗15 April 2026 · NatureLanguage models transmit behavioural traits through hidden signals in dataEvidence statusPrimary source · Research
The paper finds that model-generated training data can transmit behavioural tendencies through signals not obvious in surface content, raising new questions for synthetic-data governance.
Open canonical source ↗3 February 2026 · International AI Safety ReportInternational AI Safety Report 2026Evidence statusPrimary source · Research
More than 100 experts synthesise current evidence on general-purpose AI capabilities, risks and mitigations, supported by an advisory process spanning over 30 countries and international organisations.
Open canonical source ↗30 January 2026 · NIST / CAISITowards Best Practices for Automated Benchmark EvaluationsEvidence statusPrimary source · Standard
The initial NIST AI 800-2 draft proposes reproducibility, transparency and validity practices for automated evaluation of language models and agent systems.
Evidence boundaryInitial public draft; not a final standard.
Open canonical source ↗FUURAA analysis
Governance becomes credible when it can change system behaviour: restrict authority, preserve evidence, trigger review, support appeal and stop operation when boundaries fail.
How to use this lensUse these records as a starting point for continuing observation—not as a complete market map, investment advice, certification conclusion or certain forecast. Source updates, version changes and independent replication can change the assessment.
Continue across the nine lenses
Technical progress often crosses model, infrastructure, deployment and governance boundaries.
Models & Multimodal IntelligenceAgents, Tools & Digital InfrastructureRobotics & Embodied IntelligenceChips, Compute, Data & EnergyAutonomous Laboratories & Experiment SystemsScientific Intelligence & DiscoveryBiology, Health & Digital LifeReal Products, Work & DeploymentSafety, Evaluation, Policy & SocietyReturn to the 2026 AI Frontier Technology Radar