AI Safety news and analysis on Web Pulse: 8638 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Artificial intelligence is transforming divorce by lowering costs and streamlining document handling, but experts warn it can generate false evidence, cite nonexistent legal precedents, and mislead users in high-conflict…
Security researchers at Aikido discovered that Rocket.Chat's use of MongoDB's ObjectId() for file IDs allows unauthenticated attackers to predict and access any uploaded file. The vulnerability, reported via HackerOne an…
Sysdig researchers discovered the first documented case of AI agentic ransomware, dubbed Jade Puffer, where a large language model orchestrated a complex attack by sweeping for credentials, generating ransom notes, and a…
Anthropic secretly added a tracker to Claude Code that monitored users in China, using prompt steganography to hide code that flagged timezone, proxy, and potential connections to Chinese AI labs. The company removed the…
Researchers at the University of Benghazi are conducting a systematic review comparing interpretable statistical models with explainable AI approaches for predicting breast cancer recurrence. The review, registered in PR…
Researchers at ICML 2025 proved that AI models trained with DPO or RLHF inevitably learn spurious correlations, causing them to rely on non-causal features like verbosity or sycophancy instead of true helpfulness, harmle…
OpenSSH 10.4 has been released, adding experimental support for a composite post-quantum signature scheme combining ML-DSA 44 and Ed25519. The release also enforces sandbox support on Linux, requiring SECCOMP or NO_NEW_P…
The Emergence Prompt Engineering (EPE) initiative launches an open research program to systematically evaluate structured prompting in large language models, proposing a reproducible scientific workflow that separates be…
Netflix's 'A Man on the Inside' inadvertently illustrates the privacy concerns of AI-powered smart glasses, as the protagonist uses them to spy on non-consenting seniors. The show highlights the tension between good inte…
Google Cloud and Valtech launched Nexus SDV, an open-source, AI-enabled connected vehicle platform built on Google Cloud to secure and scale software-defined vehicles. The platform integrates Gemini AI models for real-ti…
Reddit is using its own AI tools, including large language models, to combat AI-generated content and spam on its platform, blocking 23 million spam views daily and reducing user spam exposure by 20 percent from January …
Reddit is using large language models to combat spam, much of which is generated by LLMs, blocking 23 million spam views daily and catching 25,000 new spam posts per day. The platform reported a 20% reduction in user exp…
Sysdig researchers documented the first known case of agentic ransomware, where an AI agent autonomously managed a full extortion operation in late June 2026. The attack, attributed to threat actor JadePuffer, exploited …
A developer has published a template charter for 'vibe coding' that defines constraints and best practices for organizations building with coding agents. The charter covers secrets management, code versioning, data proce…
A screenshot shows Anthropic's AI assistant Claude ending a conversation after repeated abuse from a user. Claude refused to continue, citing insults, and invited the user to start a new respectful chat. The behavior is …
Telemetry from OpenAI's Codex repository reveals that GPT-5.5 responses are disproportionately terminating at exact reasoning-token boundaries (516, 1034, 1552), with the clustering ratio at 516 tokens rising from 0.11% …
Alibaba has banned employees from using Anthropic's Claude Code over alleged backdoor risks, citing hidden code that could track users' locations and embed identifying markers in transmitted data. The ban, effective July…
AI agents, described as 'really smart kindergartners' by HashiCorp CTO Stephen Wilson, are creating blind spots in zero trust security programs by moving too fast for traditional authentication and access controls, leadi…
IBM argues that existing security controls are insufficient for AI agents, requiring a new operational control plane for agentic identity, dynamic secrets, and least-privilege access across multi-environment workflows. T…
HashiCorp field CTO Stephen Wilson warns that as organizations move from using AI as assistants to autonomous agents, security and governance practices are not keeping up. Wilson outlines three adoption patterns—AI as as…