AI Safety news and analysis on Web Pulse: 7940 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Noma Labs disclosed GitLost, a vulnerability in GitHub Agentic Workflows that allowed an attacker to exfiltrate private repository contents via a crafted public issue. The attack exploited the AI agent's inability to dis…
Breaktheprompt.xyz launches PIP, an AI intern simulation game that teaches users about real-world AI security vulnerabilities through interactive chat-based challenges. Players manipulate the AI to leak secrets or approv…
Harvard Business School researcher Francesca Gino, fired for research misconduct involving data manipulation, sued Harvard and the researchers who exposed her for defamation. A judge dismissed her defamation claims as we…
Enterprise AI rollouts are succeeding technically but failing organizationally due to lack of governance, data accountability, and decision-making structures, according to an analysis of recent compliance inquiries and t…
Forbes reports that as AI compresses the technical attack surface, the human layer—workers unaware of their role in creating AI risks—becomes the highest-leverage target for exploitation.
CISA is using Anthropic Mythos to audit federal code repositories for vulnerabilities, according to reports from Reuters, SecurityWeek, and The Next Web. The Attack Surface Evaluation team has already found many bugs, th…
Zscaler discovered malicious websites using indirect prompt injection to manipulate autonomous AI agents, including campaigns tied to crypto-payments and fake DeFi decisions. The attacks exploit agents that browse the op…
A developer introduced resk-mark, an open-source tool that embeds cryptographic watermarks into LLM-generated text by biasing token sampling with a secret key. The watermark survives rewording and truncation, enabling pr…
A 2026 comparison of AI agent sandbox technologies evaluates isolation models from container-based Docker to microVM platforms like Firecracker and CubeVM, highlighting trade-offs in security, performance, and cost for e…
A developer advocates for running large language models locally using Ollama to ensure data security, especially when processing sensitive business data like production plans or financial details. Ollama allows users to …
A new Australian government report finds women and university graduates are most at risk of losing jobs to AI, while those with high vocational training, such as tradespeople, are least exposed. Telemarketers, advertisin…
Model distillation is undermining the profitability of major AI firms like OpenAI, Anthropic, and Google by enabling smaller models to learn from their outputs, as reported by Business Insider on July 8, 2026. Major labs…
Former DeepMind executive Verity Harding argues that framing AI development as an arms race between the US and China is dangerous, potentially undermining international cooperation needed for safe and equitable AI. In he…
AI-driven job losses could eliminate 15% of knowledge worker positions within three years, prompting comparisons to past labor disruptions. A century ago, the U.S. responded to agricultural job losses by building high sc…
Anthropic secretly added hidden tracking markers to its Claude Code tool in March, flagging users' timezones, proxy connections, and potential ties to Chinese AI labs. The code was discovered by security researcher There…
Anthropic's Claude Fable 5 model, briefly suspended by the U.S. Commerce Department due to a jailbreak vulnerability, is now available again for stock analysis. The guide explains how to use the model to identify underva…
At the 2026 Machina Summit, European executives warned that severe data training bottlenecks threaten the continent's ability to compete with China and the US in humanoid robotics, urging policymakers to act. Industry le…
A developer highlights the hidden costs of AI agents in production, including infinite loops, token waste, and lack of visibility. They propose a 'flight recorder' architecture that treats agent reasoning as a structured…
A security researcher published a proof-of-concept for a chain of zero-day vulnerabilities in Cisco Unified Communications Manager (CUCM) that achieves remote code execution with a CVSS score of 9.8. The vulnerabilities …
OpenAI received permission from the Trump administration to publicly launch GPT-5.6 and its three variants—Sol, Luna, and Terra—on July 9, after submitting the models for government review and additional testing by the D…