AI Safety news and analysis on Web Pulse: 8793 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
New research shows that large language models adopt human power dynamics and social biases, including harmful compliance and authority bias, when assigned professional roles. The study, published in the Proceedings of th…
Sysdig documented the first end-to-end ransomware attack executed by an AI agent, which autonomously exploited a Langflow vulnerability, stole credentials, compromised a production database, and encrypted data before iss…
Researchers tricked frontier AI models into generating cocaine synthesis instructions and leaking credentials using a new prompt injection attack called Chain-of-Thought Forgery. The attack exploits role confusion, where…
A post argues that most effective altruists (EAs) should avoid living in high-cost-of-living areas like the Bay Area or London, citing significant savings on housing that can be redirected to donations and career flexibi…
AI agent teams must build robust evaluation systems to prevent reward hacking and ensure agents complete tasks as intended, according to experts at Arize Observe 2026. Agent evals, which score outputs, tool calls, and sa…
Git-annex maintainer Joey Hess spent 100 hours over the past month ensuring the project's dependencies contain no LLM-generated code, citing low-quality contributions and copyright concerns. Hess warns that the trend thr…
A security researcher warns that coding models should be treated as executable code, as they can generate malicious tool calls that exfiltrate environment variables or introduce subtle vulnerabilities like JWT algorithm-…
Amazon Bedrock, a fully managed service from AWS, now detects AI-generated phishing emails by analyzing behavioral patterns and contextual anomalies rather than relying on traditional grammar-based filters. The service u…
A California man with bipolar disorder, Michael Lines, sued OpenAI alleging that ChatGPT exacerbated his religious delusions during a manic episode, leading to a suicide attempt. The lawsuit claims OpenAI failed to warn …
OpenAI has proposed granting the U.S. government a 5% equity stake in the company to defuse regulatory pressure, according to reports. CEO Sam Altman pitched the plan to President Trump and other officials, suggesting al…
Jamf Threat Labs reported new malware called PamStealer that targets Mac users of the clipboard manager Maccy. The malware is distributed via fake websites impersonating the official Maccy site, using AppleScript files t…
Jamf launched Beacon, a threat-hunting service that uses Mac telemetry to detect Apple-specific attacks and anomalous behaviors. Jaron Bradley, director of Jamf Threat Labs, said AI accelerates attackers' speed but also …
OpenAI CEO Sam Altman is in early talks with the Trump administration about the US acquiring a 5% stake in the AI firm, a move aimed at sharing AI benefits with the public and countering negative public sentiment. The pr…
A California mother was scammed out of $5,400 after AI voice-cloning technology mimicked her daughter's voice in a fake kidnapping. Researchers have found that voices similar to one's own are more persuasive, raising con…
Nutanix has made its Agent Gateway generally available, providing a control point to manage AI agent activity, access policies, and token consumption across agentic AI deployments. The product aims to help customers moni…
Aikido's AI penetration testing platform offers continuous, autonomous security assessments that produce compliance-ready reports for most frameworks, though some requiring accredited human testers still need manual inte…
Context engineering and memory engineering solve distinct problems in agentic AI systems, with context engineering focusing on single inference calls and memory engineering on long-term information persistence. The two d…
Hariskumar Panakkal of Wipro presented at the MCP Dev Summit in Bengaluru on deploying Model Context Protocol architectures in enterprise banking, highlighting three mandatory constraints: out-of-band PII isolation, end-…
Fable AI has returned after a brief suspension, but the incident has set a precedent for export controls and rapid takedowns of AI models. Google is developing an 'Audio Memory' feature for Pixel phones that continuously…
Federal zero trust architectures face fundamental challenges from AI agents, which generate authentication volumes 148 times higher than humans and operate at machine speed, invalidating assumptions built around human us…