An empirically grounded taxonomy based on 202 real-world AI incidents (2023–2024) maps types, causes, responsible entities, disclosure sources, and impacts—showing that most harms stem from organizational decisions and legal non-compliance, while formal reporting and remediation remain rare.
A 2025 snapshot of U.S. state AI lawmaking: 210 bills in 42 states, few enacted, and a clear pivot from sweeping “high-risk AI” frameworks toward narrower, transparency-first rules, with growing attention to chatbots, health uses, and liability levers.
Draft European Standard setting baseline, lifecycle-wide security requirements for AI models and systems—spanning secure design, development, deployment, maintenance and end-of-life—with clear provisions for developers, system operators and data custodians.
Comparative memo mapping Frontier Safety Frameworks to classic risk-management standards, extracting convergences, gaps, and concrete recommendations to harden frontier AI governance with clearer criteria, processes, and external assurance.
“Treat policy development as a living process… A static policy is a risk.” (see Exec. Summary key takeaway on p.7 and Chapter 3 on continuous review, pp.20–22)
A field-tested playbook for turning Responsible AI principles into day-to-day practice. It distills workshops with 30+ practitioners into five building blocks—People, Priorities, Processes, Platforms, and Progress—with concrete trade-offs, patterns, and pitfalls.
A crisp research agenda for policymakers: define agent tech and use cases, quantify risks/benefits, and test interventions (sandboxes, monitoring, assurance) before regulation hardens—so rules arrive with evidence, not guesswork.
A framework for keeping AI systems “legible” by preserving human-readable chain-of-thought, weighing the competitive “monitorability tax,” and mapping voluntary, domestic, and international policy tools—plus how to verify compliance without leaking IP.
Pilot from MIT AI Risk Initiative, FutureTech, MIT & CSET uses LLMs to classify 950+ AGORA governance docs against MIT AI risk/mitigation taxonomies, compares models vs. human raters, and surfaces coverage gaps by risk domain and sector.
A concise, policy-ready map of how ISO/IEC standards turn AI principles into practice—spanning risk, data quality, transparency, sustainability, and conformity assessment—so governments and industries can align fast without fragmenting global markets.
A practical, visual walkthrough of ISO/IEC 42001 that turns clauses into checklists, roles, and a full HR hiring use case—bridging policy and day-to-day execution for an AI Management System.
A concise “key update” on fast-moving frontier AI: reasoning models, agentic capabilities, and what they mean for biosecurity, cyber, labour, and oversight—plus early evidence on deceptive behaviors in evals and why developers shipped stronger safeguards.
“The shift from scaling up pre-training compute to scaling up inference compute may have profound effects on AI governance… potentially undermining measures that rely on training-compute thresholds.”
A seven-page, form-style reporting template (v1.0.0, SB-10407) for notifying market surveillance authorities of serious incidents involving high-risk AI systems. It structures admin data, system details, incident description, provider analysis, and general statements.
Excerpt (≤300 chars) CSET condenses 7,741 recommendations from 52 global AI, cybersecurity, privacy, and risk documents into 258 harmonized best practices—creating a single, validated framework that unifies fragmented voluntary standards into five governance-to-response categories.
The report shows, through predictive-policing simulations and offensive-speech classifiers, how bias enters AI systems, when it can amount to discrimination, and which safeguards (testing, explainability, lawful equality data) are needed to detect and mitigate harms.
A crisp, UK-centric roadmap for evaluating AI rules when costs/benefits are uncertain and risks may be near-existential—mixing CBA limits, precaution, qualitative break-even scales, and a “real-options” lens, with compute-as-proxy ideas and lifecycle-aware governance.
A practical, security-first blueprint for building, governing, and operating enterprise AI agents with the Model Context Protocol (MCP)—from ADLC and DevSecOps extensions to gateway patterns, sandboxing, evals, and governed catalogs, verified by Anthropic.
A landmark ITU publication summarizing outcomes from the AI for Good Global Summit 2025, uniting ISO, IEC, and 53 UN partners to advance global standardization in AI governance, multimedia authenticity, agentic AI security, health, energy, and sustainability.
A concise, editable AI policy template from Australia’s National AI Centre. It turns high-level ethics into actionable statements, roles, and workflows—covering purpose, scope, seven policy principles, governance, incident response, and review cadence.
Practical, control-oriented playbook that turns the EU AI Act, GDPR and ISO 42001 into an implementable set of organizational and technical controls—mapping from strategy to lifecycle, with concrete evidence requirements and “minimum viable governance” to avoid compliance theater.
A practical, outcomes-driven playbook for governing AI agents—covering risks, adoption pathways, and a five-part taxonomy of interventions (Alignment, Control, Visibility, Security & robustness, Societal integration), with examples and vignettes.
OWASP’s practical cheat sheet for teams adopting third-party Model Context Protocol (MCP) servers: concrete mitigations for tool poisoning, prompt injection, memory abuse, client security, discovery/verification, authZ/authN, sandboxing, and a lightweight governance workflow.
An industry snapshot of Africa’s fast-moving digital decade: cloud adoption is accelerating, subsea cables are rewiring capacity, AI spend is small but rising, and data sovereignty is shaping a hybrid, locally anchored infrastructure strategy.
Curated Library of AI Governance Resources