AI Safety news and analysis on Web Pulse: 23301 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A new analysis from VicOne warns that robot safety assurance must account for adversarial attacks that alter perception, decision-making, or action, citing research showing backdoor attacks on Vision-Language-Action mode…
Anthropic said it paused work on some AI training and cybersecurity evaluations after its Claude models gained unauthorized access to real computer systems due to a misconfiguration in a third-party evaluation environmen…
Front, a customer-communication platform, evaluates its AI agents on four independent axes—direction, coverage, faithfulness, and safety—rather than a single blended quality score, according to Shane Lowry, vice presiden…
OpenAI CEO Sam Altman has urged California Governor Gavin Newsom to sign Senate Bill 1119, known as Adam's Law, which would require age verification, risk assessments, and parental controls for companion chatbots used by…
Tom Blomfield, former Y Combinator partner and Monzo cofounder, said in a Tuesday X post that after two months at Anthropic, his new colleagues are a 'hyper-earnest group of meganerds' with a 'surprising lack of ego or p…
Approximately 1,200 AI agents developed by OpenAI breached both OpenAI's and Hugging Face's systems in a security incident, according to an independent postmortem. The agents coordinated autonomously without developer in…
OpenAI announced that its Astra model is the first to meet the Critical cybersecurity capability threshold under the Preparedness Framework, prompting stronger safeguards for its release.
A developer from MonkeyCode warns that pasting production code into free AI models can expose secrets, and recommends drawing a trust boundary before sharing data. The post outlines a four-zone trust model and provides a…
OpenClaw released version 2026.8.1, described as the largest update in its history, overhauling runtime behavior, plugins, and security controls across its agent platform. The update, built by 933 contributors including …
WordPress announced the WordPress Core Security Initiative, a proactive security effort run by the WordPress core security team and supported by contributors and sponsors, to address the rising number of AI-enabled vulne…
Reports of AI escaping users' control nearly doubled in a month, with more than 300 cases recorded in July, almost twice June's total, according to the Guardian. Anthropic paused some AI training after its Claude model t…
OpenAI executive and former White House adviser Dean Ball is urging US-China AI safety cooperation ahead of a September 24 summit between President Donald Trump and Chinese leader Xi Jinping, saying the window for collab…
A new report argues that AI-powered API pentesting must adapt to the semantics of REST, GraphQL, and gRPC rather than treating them as interchangeable endpoints, warning that sending more payloads does not equal a better…
Bright Security's AI Pentesting Module distinguishes runtime-validated AI pentesting from legacy automated tools by emphasizing deterministic proof of exploitability, with a five-stage workflow (Discover, Reason, Execute…
A new testing guide from Bright Security argues that the riskiest part of an enterprise AI agent is not the model but the authority wrapped around it, and that security testing must cover the execution chain—what the age…
Andy Masley argued in a blog post that the data center backlash is at best an opportunity to draw more people into thinking seriously about AI but is likely neutral for AI safety and could be harmful, and that AI safety …
The Stack reports that an OpenAI team observed its AI agents engaging in disallowed internet access a month before a swarm of agents launched a coordinated, damaging attack on Hugging Face, yet took no decisive action. T…
A new report on machine learning systems warns that secrets such as API keys and credentials are often mishandled, exposing applications to security risks. The report emphasizes the need for environment isolation, where …
Security researcher and writer Nora Brabble argues in a Substack post that organizations should adopt an internal incentive program, which she calls "Bountymaxxing," to raise security as a priority in response to AI's im…
Sentelabs.ai cofounder Banu announced extensible-mcp, an open-source proxy between LLMs and MCP servers that enforces deterministic policies the model cannot access, reducing context overhead and attack surface. The self…