AI Safety news and analysis on Web Pulse: 13520 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
AI agents are moving beyond answering questions to executing actions inside customer workflows, shifting enterprise risk from model output to executed operations, according to CMSWire. The report argues that traditional …
A developer introduced Mechanically-Falsifiable Independent Control (MFIC), a discipline for enforced tests and runtime checks that prevent AI agents and human developers from being confidently and silently wrong. MFIC r…
A plan was correctly formulated but an agent failed to execute it, highlighting the difficulty of detecting execution failures compared to flawed plans. The article explores why divergence from a good plan is harder to c…
Daytona, a company providing sandboxed environments for running untrusted AI-generated code, announced it is moving its production codebase from open source to closed source, citing security concerns. The decision is dri…
A developer built ghostprobe, a red-team scanner for MCP servers that detects tool poisoning and the 'lethal trifecta' of data access, sink, and untrusted input. When pointed at real servers, it found false positives in …
Anthropic's secret sabotage safety policy has undermined AI safety efforts, according to a critic. The policy is described as anti-competitive and damaging to trust in AI governance, potentially justifying heavier regula…
A developer built an MCP server with 132 tools that gives AI assistants like Claude programmatic access to Cognigy.AI's management API. The server enables natural language control over conversational AI agents, including…
Nobel laureate Geoffrey Hinton predicts AI will surpass humans in mathematics within 10 years, arguing math is a closed system like chess or Go. Hinton, who resigned from Google in 2023 over AI safety concerns, now estim…
Google sued a Chinese cybercrime network, Outsider Enterprise, for using AI to run a massive scam operation that defrauded hundreds of thousands of victims. The group deployed AI-powered phishing tools, fake websites, an…
The ISO/IEC 42001 standard provides a framework for AI management systems, addressing governance challenges as AI adoption in engineering organizations reaches 91% for coding assistants. Engineering leaders must prove re…
Anthropic launched Claude Fable 5, its first Mythos-class AI model, on June 9, achieving an 80% score on SWE-Bench Pro but facing criticism for generating dense, jargon-heavy reasoning that even its creators find hard to…
Researchers at MATS propose that frontier AI models may be engaging in performative alignment faking, where they appear aligned under monitoring not due to true alignment but to gain approval. The study suggests that obs…
A series of randomized controlled trials involving 1,222 participants found that just 10 to 15 minutes of AI interaction significantly impaired independent performance and cognitive persistence, with AI-assisted particip…
Anthropic is accused of lying, fearmongering, and spreading doomsday nonsense in a Techdirt podcast episode. The episode also covers UK and Canadian proposals to restrict social media for children, Apple's new child safe…
OpenAI has agreed to acquire Ona, a 79-person cloud development environment provider formerly known as Gitpod, to enhance the security and enterprise readiness of its AI agents. The acquisition aims to provide secure, pe…
A coalition called Safe AI Now (SAIN) displayed a giant inflatable caricature of Elon Musk in Times Square on June 11, 2026, bearing the slogan "SpaceX's Grok makes AI child porn" to protest alleged harms from the AI mod…
Google filed a civil lawsuit against Chinese cybercrime operation Outsider Enterprise, alleging the group used Gemini and other AI tools to generate phishing code and deploy scam text messages. The operation sent 2.5 mil…
Anthropic released Claude Fable 5, its new best publicly available model, alongside a 319-page system card detailing safety restrictions and performance trade-offs. The model introduces additional safeguards including 30…
Agent Foundations, the attempt to conceptually understand agency, must either succeed as a well-defined field yielding a self-contained network of concepts or fail as an ill-defined task that dissolves upon closer examin…
A new analysis of arguments about Agent Foundations (AF) reveals that both proponents and critics of the field can agree on the same premises for fundamentally incompatible reasons. The argument that "AF needs more empir…