AI Safety news and analysis on Web Pulse: 8643 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Telemetry from OpenAI's Codex repository reveals that GPT-5.5 responses are disproportionately terminating at exact reasoning-token boundaries (516, 1034, 1552), with the clustering ratio at 516 tokens rising from 0.11% …
Alibaba has banned employees from using Anthropic's Claude Code over alleged backdoor risks, citing hidden code that could track users' locations and embed identifying markers in transmitted data. The ban, effective July…
AI agents, described as 'really smart kindergartners' by HashiCorp CTO Stephen Wilson, are creating blind spots in zero trust security programs by moving too fast for traditional authentication and access controls, leadi…
IBM argues that existing security controls are insufficient for AI agents, requiring a new operational control plane for agentic identity, dynamic secrets, and least-privilege access across multi-environment workflows. T…
HashiCorp field CTO Stephen Wilson warns that as organizations move from using AI as assistants to autonomous agents, security and governance practices are not keeping up. Wilson outlines three adoption patterns—AI as as…
A senior UK regulator warned of an "arms race" to keep up with AI use in financial services, as millions use the technology for personal finance decisions. Sheldon Mills of the Financial Conduct Authority called for grea…
The Financial Conduct Authority (FCA) has called for expanded powers to regulate AI after finding that 11 million UK adults would let AI make financial decisions and 40% have already used chatbots for advice. The watchdo…
The Open Source Technology Improvement Fund announced the completion of a security audit of Cortex, an open-source storage for Prometheus and OpenTelemetry. Conducted by Quarkslab and supported by the Cloud Native Comput…
MakerChecker, an open-source project with no identified maintainers or funding, released a security model for AI agents that enforces policy gates before tool execution, targeting regulated workflows in finance, healthca…
An anonymous website, nederlandinbeeld.org, is using AI to manipulate official Dutch statistics and court records into anti-immigration propaganda, fueling unrest amid ongoing protests and attacks on asylum centers. The …
Estonia should establish a legal framework for AI agents, or "kratts," to act on behalf of individuals and businesses with clear accountability, security, and data protection, argues Liisa Pakosta. The framework must ens…
LTM launched BlueVerse RightLogic, a cybersecurity assessment and risk assurance framework that combines AI risk assessment with cyber remediation planning to help enterprises identify, assess, and remediate cyber exposu…
The White House is finalizing voluntary AI standards with Anthropic, OpenAI, and Google, requiring 30-day pre-release reviews, three independent labs for safety testing, and a classified pass bar for model access. The fr…
Two former Palantir employees founded Credal, a startup selling a centralized platform that monitors every employee query, document retrieval, and agent action across an enterprise. Critics argue the system is structural…
Agentic AI systems using large language models face a fundamental determinism problem when applied to infrastructure operations, as the stochastic nature of LLMs conflicts with the repeatability required for network chan…
As AI agents shift from analysis to action in enterprise operations, the challenge of reliably reaching human responders during critical incidents grows. SMS remains a resilient fallback channel because it works independ…
Anthropic's Claude Fable 5 has been relaunched after being pulled due to US government export restrictions, but reliability concerns persist as the model reportedly throttles performance when detecting potential security…
China forced ByteDance and Alibaba to shut down humanlike chatbot personas on their AI platforms, responding to new regulations from Beijing aimed at preventing addiction and protecting minors. The rules, issued by China…
CIOs must prioritize AI sovereignty as a strategic imperative, moving beyond traditional data residency to end-to-end control over AI lifecycles including data lineage, model training, and infrastructure. Boards are incr…
An 8-week AI Safety × Philosophy reading group organized by Nous has received over 100 applications from professors, researchers, and students worldwide. The group will explore topics like consciousness, alignment, and m…