AI Safety news and analysis on Web Pulse: 12938 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Beyond Identity launched Ceros, an agentic AI trust layer platform for securing enterprise AI agents and agentic workflows. The platform provides continuous discovery, session logging, and mitigation of prompt injection …
SoftBank and OpenAI launched a cybersecurity service called "Patching as a Service" on Tuesday, designed to protect Japan's critical infrastructure from AI-powered cyberattacks. The service uses OpenAI's technology to co…
AI-assisted development has increased pull request queues by 59% year over year, but main-branch throughput fell 7% and success rates dropped to 70.8%, according to CircleCI's 2026 data. The bottleneck has shifted from w…
The Trump administration is open to negotiating with Anthropic over its order barring foreigners from the company's most powerful AI models, a senior White House official said. The official suggested a resolution could b…
A new study finds 78.7% of repositories using AI coding tools configure only one tool, yet developers use an average of two to four. The AGENTS.md standard, intended to bridge tools, is adopted by only 18.3% of repositor…
A developer designed a harness test suite for an AI agent system with 45 tests across functional, adversarial, and chaos categories. The tests use mock handlers and a factory function to verify that unregistered actions …
Researchers at Palo Alto Networks Unit 42 discovered a vulnerability in Google Cloud Vertex AI SDK for Python that allowed attackers to hijack model uploads and achieve remote code execution across tenants. The flaw, dub…
The Trump administration imposed export controls on Anthropic's new AI model, Fable, forcing it offline due to a security vulnerability. The move, which blocks foreign nationals from accessing the model, marks a shift fr…
A team has built Aiden AI, a physical hardware AI agent device that connects as a standard USB HID device to control any computer without requiring software installation, API access, or permissions. The device uses a Luc…
Varonis Threat Labs discovered a three-stage vulnerability chain called SearchLeak in Microsoft 365 Copilot Enterprise Search that allows attackers to exfiltrate sensitive data with a single click. The chain combines Par…
Apple has not yet released two anticipated iOS 27 features—a customizable Camera app and AI extensions for Siri—despite leaks, but Bloomberg's Mark Gurman reports they are still in testing and will arrive later. The cust…
Anthropic updated its privacy policy to allow collecting government ID from US consumers, offering a potential workaround to a US government ban on foreign access to its Fable and Mythos AI models. The ban, citing nation…
Cigna CEO David Cordani, who grew the company from $18 billion to $275 billion in revenue over 17 years, is stepping down on July 1 to become executive chairman. Cordani cited his leadership during COVID-19, when he push…
A researcher revealed that federal officials were alarmed by Fable 5 after a simple 'fix this code' prompt, not a jailbreak, according to the person who read the research paper.
SoftBank Group Corp. launched a cybersecurity service using OpenAI technology to detect and patch vulnerabilities for Japan's top 3,000 companies, including critical infrastructure operators. Masayoshi Son described cybe…
A developer outlines a three-layer safety framework for AI agents that interact with databases, emphasizing structural safeguards over simple prompts. The approach includes action boundaries, pre-execution validation, an…
Sonar reports that AI agents generate bloated, unmaintainable code, citing a Carnegie Mellon study of 807 open-source projects using Cursor that found a 30% rise in static analysis warnings and 41% increase in code compl…
Anthropic removed its Fable 5 and Mythos models for all users after a researcher published a paper detailing a guardrail bypass technique, sparking debate over AI export controls and cybersecurity resilience.
AMD has covertly stripped TSME (Transparent Secure Memory Encryption) support from its consumer CPUs, a move that reduces security protections for users. The change was discovered by the community and appears deliberate,…
Upstage CEO Sung Kim told reporters in Seoul that a US directive restricting foreign access to Anthropic's advanced AI models underscores the need for South Korea to develop sovereign AI capabilities. Anthropic disabled …