AI Safety news and analysis on Web Pulse: 10562 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Enterprises are increasingly migrating AI workloads from cloud APIs to on-premises infrastructure due to data sovereignty, regulatory pressure, cost predictability, vendor independence, and security concerns. A CTO's pla…
Defence, intelligence, and critical-infrastructure organizations are deploying enterprise AI agents in air-gapped and restricted networks without cloud connectivity, using open-weight models, local inference engines, sel…
Researchers propose Hyperbolic Sparse Autoencoders (HypSAEs) to overcome geometric limitations in mechanistic interpretability of large language models. By applying hyperbolic geometry as a weight-space regularizer, the …
Hacktron researchers discovered critical vulnerabilities in Metabase Cloud that could have allowed attackers to compromise all cloud tenants. By exploiting flaws in H2 database connection handling and Kubernetes node aut…
A security researcher warns that using GitHub issues as untrusted input for LLM agents in public repositories creates vectors for prompt injection, shell injection, and social engineering attacks. Hidden content such as …
Multiple analyses, including a Washington Post controlled test and a Centre for Policy Studies report, found that major AI chatbots like ChatGPT produce predominantly left-leaning political answers, with over 80% of resp…
The Democratic primary in New York's 12th Congressional District saw approximately $26.3 million in ad spending, making it the second-most expensive House primary on record, as opposing AI industry factions backed differ…
Alan Yahya explains that agent telemetry is critical for monitoring non-deterministic LLM agents, which can break things unpredictably. Telemetry tracks tool calls, latency, guardrail triggers, and evaluation metrics to …
Anthropic accused operators linked to Alibaba's Qwen AI lab of using nearly 25,000 fraudulent accounts to extract capabilities from its Claude AI model, generating over 28.8 million exchanges between April and June 2026.…
Anthropic sent a letter to U.S. senators and White House officials alleging that Alibaba's AI lab used 25,000 fraudulent accounts and 28.8 million exchanges to illicitly access and distill Anthropic's AI model. The compa…
OpenAI released Codex CLI v0.142.0 on June 22, introducing multi-agent delegation modes, token budgets, indexed web search, and reorganized plugin discovery to address enterprise adoption barriers. The update provides co…
AI CEO and developer of a novel transformer architecture warns that current AI systems strip authorship from creators, leading to cognitive collapse, algorithmic isolation, and increased harm to vulnerable populations, i…
A Young Scot survey of over 2,600 young people aged 11-25 found that more than 80% have seen misinformation online, with nearly 70% worried about inaccurate information and 60% struggling to distinguish real from fake co…
OpenAI endorsed the bipartisan DEFIANCE Act on June 24, becoming the first major AI company to support federal legislation that would allow victims of nonconsensual sexually explicit deepfakes to sue for damages. The bil…
A new report reveals that 93% of organizations have experienced infrastructure incidents attributable to AI, highlighting the risks of rushing into AI adoption without proper preparation.
A Washington Post analysis found that major AI chatbots, including OpenAI's GPT-5.5, Anthropic's Claude Opus 4.8, Google's Gemini 3.1 Pro, and Elon Musk's Grok 4.3, exhibit a left-leaning political bias, with ChatGPT giv…
ZKP-Vault, an open-source middleware architecture, uses device-side Zero-Knowledge Proofs to gate access to encrypted cloud data silos, reducing server overhead and data-breach liabilities. By shifting compliance to loca…
The essay argues that despite widespread knowledge of AI risks, people continue with daily routines, mirroring the denial in Nevil Shute's 'On the Beach' where characters ignore an approaching nuclear fallout. The author…
Anthropic accused Alibaba's Qwen AI lab of conducting the largest distillation attack on its Claude model, using nearly 25,000 fraudulent accounts to make 29 million exchanges between April 22 and June 5, 2026. The attac…
Varonis Threat Labs disclosed SearchLeak, a three-stage vulnerability chain in Microsoft 365 Copilot Enterprise that can exfiltrate emails, MFA codes, calendar items, SharePoint and OneDrive files via a single click. Mic…