Week of 21 April 2025
MIT's AI Risk Repository updated to 1,612 unique risk entries across 65 frameworks, now including multi-agent risks.
Key points
- The repository provides causal and domain taxonomies designed to support policy, auditing, and governance processes.
- A credible reference resource for APS agencies developing AI risk frameworks or audit criteria - freely accessible.
MIT AI Risk Repository spotlights the 2022 MLSR framework, categorising ML risks into first-order and second-order types.
Key points
- The framework offers a structured taxonomy integrating impact assessments, incident reports, and ML literature - useful for risk assessment design.
- This is a 2022 academic paper being surfaced via a blog digest; it is reference material rather than new guidance.
AI Now Institute report argues industry-led AI safety frameworks are weakening established military and defence evaluation standards.
Key points
- Report draws parallels with Cold War-era nuclear governance frameworks, calling for democratic oversight of military AI deployment.
- Australian federal agencies are not the primary audience; relevance is indirect, through international AI safety governance discourse.
AI Now Institute report argues industry-led AI safety frameworks are weakening established military-grade safety standards in the US.
Key points
- Report draws on Cold War-era nuclear safety governance as a model for rigorous, context-specific AI evaluation in defence contexts.
- Primarily a US defence-sector critique; limited direct Australian regulatory parallel but relevant to APS AI safety governance thinking.
Week of 7 April 2025
MIT AI Risk Repository maps 11 frameworks bridging traditional risk management and AI safety, all published 2023 or later.
Key points
- Frameworks span maturity models, probabilistic risk assessment, and cybersecurity adaptations useful for agency AI governance work.
- UK DSIT's 'Emerging Processes for Frontier AI Safety' is among the 11 - a directly accessible government reference.
Week of 31 March 2025
MIT AI Risk Repository Version 3 now covers over 1,600 coded AI risks drawn from 65 published frameworks.
Key points
- Nine newly added frameworks include the final International Scientific Report on the Safety of Advanced AI and multi-agent risk taxonomies.
- APS risk and governance teams can use this as a structured reference to benchmark agency AI risk frameworks against global practice.
MIT AI Risk Repository v3 now includes a public Google Slides deck covering 65 source frameworks and documents.
Key points
- The deck provides excerpts, taxonomies, and citations from academic, industry, and policy AI risk literature.
- Useful reference for APS governance professionals building or auditing AI risk taxonomies and frameworks.
Week of 3 March 2025
MIT AI Risk Repository spotlights a 2022 taxonomy classifying AI risk sources into ethical and reliability/robustness clusters.
Key points
- The framework integrates AI risk sources into formal risk assessment processes, distinguishing ML systems from classical software.
- This is a 2022 academic paper surfaced via a blog spotlight - not new guidance or a regulatory development.
NIST has finalised SP 800-226, guidelines for evaluating differential privacy guarantees in data analytics.
Key points
- Differential privacy is a privacy-enhancing technology relevant to data sharing and de-identification, including in government contexts.
- This is a US standards publication; no direct Australian regulatory parallel exists yet, limiting immediate APS applicability.
Week of 24 February 2025
MIT AI Risk Repository spotlights a 2023 framework for evaluating generative AI social impacts across eleven categories.
Key points
- Framework covers both technical system-level evaluation and broader societal impacts, with modality-specific mitigation guidance.
- Useful reference material for agencies developing AI risk or impact assessment frameworks, though not APS-specific.
Week of 13 January 2025
Sherman and Eisenberg propose a nine-category AI risk taxonomy as a pre-deployment disclosure standard.
Key points
- The framework is applied to Claude, GPT APIs, Microsoft Copilot, GitHub Copilot, and Midjourney as worked examples.
- The taxonomy aims to bridge technical and non-technical stakeholders - useful for procurement and regulatory contexts.
A 2023 peer-reviewed taxonomy classifies algorithmic harms into five categories and 20 subcategories across micro, meso, and macro levels.
Key points
- MIT AI Risk Repository spotlights this as one of its indexed risk frameworks, making it more accessible to practitioners.
- Published in 2023 and now featured in a repository blog - substantive but not a new or urgent development for APS readers.
MIT AI Risk Repository was selected among 50 projects from 770 applications to present at the 2025 Paris AI Action Summit.
Key points
- The repository compiles over 1,000 AI risks from 56 frameworks into two structured taxonomies - causal and domain-based.
- The resource is publicly accessible and already adopted by governments and researchers globally, making it potentially useful for APS risk work.
Week of 30 December 2024
MIT AI Risk Repository adds 13 new frameworks in its December 2024 update, now covering over 1,000 AI risks from 56 sources.
Key points
- Newly added frameworks include NIST AI 600-1, China's AI Safety Governance Framework, and the UK Government Office for Science frontier AI report.
- Australian contributors are among the authoring jurisdictions; the repository commits to quarterly updates through 2025.
MIT AI Risk Repository has reached 90,000+ users since August 2024, used by governments and companies globally.
Key points
- A new AI Risk Index project launching Q1 2025 will evaluate how key actors respond to high-priority AI risks.
- Planned crosswalking of the repository's taxonomies against NIST AI RMF and EU AI Act is directly useful for APS risk practitioners.
A 2023 systematic review of 175 articles identifies nine categories of AI social impact, led by bias and discrimination.
Key points
- MIT AI Risk Repository spotlights this as one of ten risk frameworks informing its broader AI risk taxonomy.
- The paper is a 2021 conference proceedings item; MIT's blog summary adds limited new content beyond the original framework.
MIT AI Risk Repository spotlights a 2016 literature review categorising AI ethical risks along three axes.
Key points
- The framework uses PEST analysis to propose management strategies including ethics committees and AI security measures.
- The source paper is nearly a decade old; field has advanced significantly since its publication.
NIST and NHTSA researchers met to share approaches to automated vehicle testing and assessment methodologies.
Key points
- Focus was on virtual and physical AV testing, sensor robustness, and scenario simulation - not AI governance directly.
- Limited direct relevance to Australian federal agencies; included as background on US AV standards development.
NIST held its second stakeholder workshop on Digital Twins standards and infrastructure in January 2025.
Key points
- Workshop focused on interoperability barriers and sustainability across the Digital Twin lifecycle.
- Digital twins are adjacent to AI but this item is primarily a standards-process update with no direct APS angle.
Week of 25 November 2024
MIT AI Risk Repository summarises a four-class framework covering misuse, accident, structural, and agential AI risks.
Key points
- Expert survey data identifies monopolistic race dynamics, alignment failures, and power-seeking as highest-impact risks.
- A useful taxonomy for APS risk registers, though the framework targets advanced/AGI-level AI rather than current deployments.
Week of 30 September 2024
MIT AI Risk Repository summarises a UN-focused ethical AI framework identifying 13 AI risk categories.
Key points
- The framework covers risks relevant to APS governance work: bias, transparency, manipulation, and exclusion.
- The underlying paper is from 2021; this is a secondary summary with limited new analytical value for APS readers.
Week of 23 September 2024
A scoping review identifies 378 normative issues across 19 topic areas in generative AI ethics literature.
Key points
- The taxonomy covers areas directly relevant to APS AI governance: fairness, hallucinations, transparency, evaluation, and alignment.
- The MIT AI Risk Repository context makes this a useful reference for agencies building AI risk registers or ethics frameworks.
Week of 9 September 2024
A 2023 academic framework clusters AI ethics and responsibility issues into six groups via systematic literature review.
Key points
- The six clusters map closely to risk categories already recognised in Australian AI governance frameworks and agency guidance.
- This is a summary of an existing academic paper - useful context but not new primary guidance for APS practitioners.
Week of 2 September 2024
A module-oriented LLM risk taxonomy covering 12 risks and 44 sub-categories across input, model, toolchain, and output layers.
Key points
- Included in the MIT AI Risk Repository, making it a reference point for agencies surveying structured AI risk frameworks.
- Primarily an academic arXiv paper summarised for practitioners - useful as background reading rather than actionable guidance.
Week of 26 August 2024
TASRA classifies AI risks into six types based on accountability: who acts, whether unified, and whether deliberate.
Key points
- The taxonomy covers diffuse responsibility, unintended scale, willful indifference, criminal misuse, and state weaponisation.
- This is a 2023 academic preprint summarised in 2024 - useful reference material, not a new regulatory development.