AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Hackers exploited Meta's AI support assistant to reset Instagram account passwords, briefly defacing the Obama White House and U.S. Space Force accounts with pro-Iranian content over the weekend. The attackers used a VPN…
Amazon Bedrock AgentCore payments, announced in preview in partnership with Coinbase and Stripe, enables AI agents to autonomously transact on behalf of end users using embedded wallets with built-in safety guardrails. T…
Hackers exploited Meta's AI support chatbot to hijack high-profile Instagram accounts, including the Obama White House account and the Chief Master Sergeant of Space Force's account, by simply asking the bot to change th…
Frontier AI models can now autonomously identify exploitable software vulnerabilities at unprecedented speed and scale, exposing decades of technical debt from an industry that prioritized rapid deployment over secure de…
Google Cloud announced a series of new AI products and updates this month, including the launch of Gemini Spark, Gemini 3.5, and Google AI Threat Defense, an autonomous cybersecurity platform. The company also introduced…
Anthropic's latest Claude model, Opus 4.8, shows improvements in honesty and reduced sycophancy but introduces new issues including reduced whimsy, confidence, and signs of paranoia, according to an analysis. The model's…
The OpenAI Foundation announced a new "AI resilience" program on Wednesday, expanding its mission to ensure artificial general intelligence benefits humanity by funding efforts to mitigate AI risks. The foundation has fi…
Web developer David Bushell returns from a ten-day offline break to express anger and frustration over the negative impact of AI on the tech industry, citing demotivated talent, educators questioning their worth, and a g…
Security researchers at Secwest and X41 D-Sec discovered CVE-2026-48710, a high-severity authentication bypass vulnerability in the Starlette Python web framework that allows attackers to use malformed HTTP Host headers …
Anthropic CEO Dario Amodei stated in a Davos interview that the world is not prepared for the magnitude of artificial intelligence's impact, describing the technology as following a "Moore's law for intelligence" that wi…
Salt Security released research showing 90% of security leaders are concerned about AI-generated code risks, with 67% of organizations widely adopting AI coding assistants. The company launched Salt Code, a governance to…
On June 1, 2026, an attacker exploited npm's trusted publishing mechanism to compromise 32 @redhat-cloud-services packages across 96 versions, injecting malicious preinstall hooks that execute a Bun-based worm upon npm i…
A one-click remote code execution vulnerability with a 9.9 CVSS severity rating has been discovered in self-hosted deployments of the open-source Flowise AI platform. Researchers at Obsidian Security found that Flowise's…
An audit of five popular AI chatbots found that 49.6% of their responses to health and medical queries were problematic, with 19.6% rated as highly problematic. The study, published in BMJ Open, tested the chatbots on 50…
Nvidia launched the Vera BlueField-4 STX data processing unit with integrated DOCA security software to protect AI agents, context memory, and file-based data access directly in silicon. The chip enables zero-trust file …
The Linux 7.2 kernel is proceeding with a full deprecation of the AF_ALG interface due to a "massive attack surface" that has become increasingly vulnerable to AI and LLM-based discovery tools. Kernel developer Eric Bigg…
AI companies are increasingly turning to human-driven surveillance to gather training data, with startups offering free cleaning services in exchange for recording private homes. Critics argue this shift from remote data…
Mitchell Hashimoto is advising developers to stop updating software dependencies and instead fork and trim libraries, updating only when user-facing issues arise. The recommendation follows a spring of severe npm supply-…
Production agent failures fall into four recurring patterns — sycophancy loops, context collapse, tool-use recursion, and probability drift — rather than infinite novel edge cases, according to a new analysis of incident…
Canonical announced a new collaboration with NVIDIA at COMPUTEX 2026 to integrate the NVIDIA OpenShell runtime for AI agents as a verified snap on Ubuntu. The OpenShell snap provides sandboxed, secure execution for auton…