AI Safety news and analysis on Web Pulse: 11274 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A study by IIIT Hyderabad finds that government procurement contracts and tenders are shaping AI governance in India more than formal regulation, embedding requirements on standards, accountability, compliance, and perfo…
The Model Context Protocol community released Enterprise-Managed Authorization (EMA), a security update that lets organizations use their identity provider to control MCP server access, eliminating per-user OAuth flows. …
Engineering teams are discovering that expanding AI context windows degrades agent reliability due to the 'lost in the middle' problem, where models fail to attend to information in the middle of long prompts, causing si…
Canada has banned sophisticated deepfakes of political figures under Bill C-25, which amends the Canada Elections Act to prohibit creating or distributing AI-generated imitations likely to mislead voters. The law targets…
Sundial launched a collaborative document editor designed to let humans review and approve AI agent work at granular levels, addressing the growing challenge of verifying agent-generated content. The platform supports an…
Norway will ban generative AI use for elementary school students from first through seventh grade starting late August, Prime Minister Jonas Gahr Stoere announced. The policy restricts AI for teens aged 14-16 to supervis…
Homebrew 6.0 ships a Linux sandbox that uses systemd's cgroup v2 and capability bounding to restrict formula access at install/run time, limiting the blast radius of buggy or malicious code. The sandbox is not containeri…
A new open-source plugin called Hallucinating Canary provides early warnings of response degradation in Claude Code sessions by detecting context loss through planted marker tokens. The plugin monitors whether these 'can…
The Trump administration imposed export controls on Anthropic's Fable model, limiting access to U.S. citizens and forcing Anthropic to make the model unavailable. Ben Thompson argues that the incident highlights Anthropi…
Cory Doctorow argues that the AI boom is driven by bosses' desire to replace workers, not by genuine progress, and warns that the technology's hype masks power dynamics that harm users. In a podcast interview, he expands…
The Model Context Protocol's Enterprise-Managed Authorization extension, stable since June 18, 2026, redefines authorization as a context-scoping mechanism that determines which data sources an AI agent can access. By re…
Security researcher found that several Claude Code community plugins listed in Anthropic's official marketplace were vulnerable to repo-jacking, where abandoned GitHub repository names could be claimed by attackers. The …
Hackers are increasingly exploiting trusted developer tools rather than breaking into systems, with two major campaigns this week highlighting the trend. TeamPCP injected malicious code into over 1,000 open-source packag…
Diagrid Catalyst, built on Dapr, introduces verifiable execution lineage for agent workflows, addressing a gap in existing workflow engines like Temporal that lack cross-boundary trust verification. The new capability al…
Yann LeCun, a prominent AI researcher, criticized Elon Musk's xAI as a 'failure' and warned that AI labs risk causing a 'big bubble explosion.' The remarks were made in a CNBC article that is currently inaccessible due t…
John Edwards, the UK's Information Commissioner and top data and AI regulator, resigned on Friday following a workplace investigation into his conduct, marking the first resignation in the office's 40-year history. His d…
OpenTunnel, a new open-source tool, enables AI agents to run commands on remote machines via ephemeral, end-to-end encrypted tunnels without SSH, accounts, or persistent access. The tool downloads a temporary CLI, requir…
Holden Karnofsky, a prominent figure in AI safety, compiled a list of ways AI safety efforts could be net negative, acknowledging that actions intended to improve safety might inadvertently cause harm. He estimates a 60%…
Security firm Paradigm Shift disclosed an unpatchable BootROM exploit, dubbed usbliter8, affecting Apple's A12 and A13 chips in devices including iPhone XS through iPhone 11, iPhone SE 2nd gen, select iPads, Apple Watch …
New benchmarks reveal that larger AI models like GPT-5.5 and DeepSeek V4 Pro hallucinate significantly more than smaller open-weight models, with GLM-5.2 achieving a 28% hallucination rate compared to GPT-5.5's 86%. The …