Mahesh Narayan @mahesh_narayan

OpenAI’s new research examines why hallucinations persist in language models, even as capabilities improve. The analysis finds that training and evaluation practices encourage models to guess rather than acknowledge uncertainty, leading to confident but incorrect outputs. These…

The UAE’s 32B-parameter K2 Think was jailbroken hours after release by researcher Alex Polyakov, who exploited a flaw he called “partial prompt leaking.” Transparent reasoning logs showed exactly which safety rule flagged his initial jailbreak attempts, enabling him to…

Beijing has reportedly barred Chinese tech firms from purchasing Nvidia’s H20 and RTX Pro 6000D AI chips, effectively freezing the company out of the market. Analysts say the move reflects China’s growing confidence in its domestic semiconductor sector and serves as leverage in…

The UN Security Council convened its 10005th meeting to debate artificial intelligence and international peace and security. Leaders, diplomats, and experts explored AI’s dual role: risks in warfare and autonomous weapons, and opportunities in peacebuilding and sustainable…

Shadow AI is rapidly altering enterprise risk profiles. Employees bypass sanctioned systems to use external AI tools, creating exposure to data leaks, regulatory scrutiny, and ethical lapses. KPMG finds 44% of employees use AI against policy, and over half trust outputs without…

China is treating AI safety as a foundation for innovation, not a brake on progress. It has mandated pre-deployment risk checks, removed thousands of unsafe systems, and accelerated standard setting. With biosecurity, cybersecurity, and loss-of-control risks crossing borders,…

NVIDIA Research outlines a shift in agentic AI toward small language models (SLMs). While large models have been central so far, SLMs offer efficiency, lower costs, and adequate capability for repetitive, specialized tasks. Their modularity and potential for edge deployment align…

A new study, The AI Model Risk Catalog, analyzed 460,000 Hugging Face model cards and identified 2,863 unique risks. Researchers compared developer-reported risks with the MIT Risk Repository and the AI Incident Database. Developers focus on technical issues like bias and safety,…

The UN General Assembly has approved the creation of two governance mechanisms for artificial intelligence: the United Nations Independent International Scientific Panel on AI and the Global Dialogue on AI Governance. These initiatives are intended to connect research with policy…

The UN General Assembly has approved the creation of two governance mechanisms for artificial intelligence: the United Nations Independent International Scientific Panel on AI and the Global Dialogue on AI Governance. These initiatives are intended to connect research with policy…

AI interpretability is emerging as a strategic imperative for trust, safety, and accountability in advanced models. This analysis outlines two complementary approaches: mechanistic interpretability, offering precise but narrow insights, and representation interpretability,…

The MIT AI Risk Initiative’s Mapping AI Risk Mitigations report consolidates 831 measures from 13 leading frameworks into a draft taxonomy covering governance, technical safeguards, operational processes, and accountability mechanisms. While operational controls like testing,…

The CFA Institute’s latest research underscores that explainable AI (XAI) is no longer optional for financial institutions. As AI drives credit decisions, investment strategies, and fraud prevention, transparency and accountability have become strategic imperatives. The report…

IBM’s AI Risk Atlas offers a unified framework for navigating the expanding risk landscape of generative and agentic AI. By consolidating risks across data, inference, outputs, governance, and societal impacts, and aligning them with leading standards such as the NIST RMF, EU AI…

Africa’s AI landscape is at a critical inflection point. While AI promises transformative gains, the continent faces disproportionate risks—electoral manipulation, cybercrime, data dependency, labor disruption, and environmental strain. With only 26.8% of states implementing…

Unlock the future of trustworthy AI in electronics! Why attend the AI Trust in Electronics Summit? Join us live from Oslo on 4 September 2025. Register free: digital.nemko.com/summit-web #AITrustinElectronicsSummit2025 #AITrust #AIGovernance #ResponsibleAI #Electronics

RBI FREE-AI Committee Calls For Sector-Specific AI Models, Etc. - MediaNama medianama.com/2025/08/223-rb… #ResponsibleAI

Edge AI is enabling real-time intelligence across industries, from healthcare to manufacturing, yet its true potential depends on systems that are both robust and reliable. These solutions must withstand environmental stress, hardware failures, variable data quality, and evolving…

Edge AI is redefining data intelligence by enabling real-time insights, reduced latency, and stronger privacy through on-device processing. However, its distributed nature creates a complex compliance landscape—spanning data security, explainability, and governance. As regulatory…

Edge AI is redefining how intelligence is embedded in cars, homes, and factories by enabling decisions at the source of data. This proximity brings speed and efficiency, but also elevates the stakes—bias or system failures at the edge can translate into real-world harm and eroded…