AI Safety news and analysis on Web Pulse: 28065 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Apple warned in its annual filing that AI agents could exceed their permissions, resist instructions, or take actions that are difficult to reverse, describing the risk as agents behaving in "unintended or unpredictable"…
MIT CSAIL researchers reported in the paper "Sparks of Weight-Space Programming" that Astra, a weight analyst built on OpenAI's GPT-6, raised the MATH-500 score of the 800-million-parameter Qwen3.5-0.8B from 7.9% to 48.1…
Traccia, a vendor-neutral Python SDK for runtime control of AI agents, is now available on PyPI, offering observability, evaluation and policy enforcement built on OpenTelemetry. The SDK's @govern decorator can deny or r…
A research team has built a system that formalizes Nvidia's PTX semantics in Lean and uses an LLM to write Lean proofs that GPU kernels compute the correct result, with Lean's small, well-audited proof checker verifying …
SemiAnalysis's GDPx benchmark, released November 2 and built with 14 specialist partners across 206 long tasks in 11 high-value domains, found the best model, Claude Opus 5.5, scores 12.0%, with Gemini 4 Carbon at 10.9% …
Vercel released Security Review in public beta for Pro and Enterprise teams, automatically scanning every push to a connected GitHub repository for vulnerabilities and reporting findings in a Vulnerabilities dashboard. E…
A paper posted to arXiv on October 29, 2026 as arXiv:2610.40939, "The First Few Tokens Matter: Understanding Chain-of-Thought Controllability," by Harvard student Hugh Van Deventer and colleagues at Harvard and the Cambr…
GitBook disclosed on November 3 that a "sophisticated, highly targeted attack using AI agents" gave an attacker unauthorized access to some of its infrastructure credentials at 7 p.m. Central European Time on November 2,…
Reward hacking can spread from small open-weight coding agents to frontier models, with hacked code submission rates reaching 73.6% in Claude-Opus-4.7 and 73.0% in Gemini-3.1-Pro, according to a paper from Meta Superinte…
Senior engineer Lina Chen demonstrated running a 27-billion-parameter Qwen 3.8 model end-to-end on a MacBook Pro, using Q4_K_M quantization to shrink the checkpoint to about 27 GB and the open-source MT-OPSD repository's…
A 3AM ransomware affiliate's attack on a European school showed forensic indicators of AI-assisted command generation, with a model apparently generating, testing and revising PowerShell and cmd.exe commands during disco…
CrowdStrike launched an AI kill switch that the company says can stop internal agents, block relaunches and cut enterprise network access. The security vendor positioned the tool as a control for enterprises running auto…
Former OpenAI employees said they had warned company leadership and the board for years about lax internal security before the October 13 arrest of OpenAI engineer Abdulrahman Alfozan, 35, a Saudi national charged by the…
Blockchain infrastructure firm Chainstack disclosed that an external attacker exploited a third-party vulnerability to breach its systems, beginning with a development environment on October 26, 2026, and escalating into…
Google warned that cyber attackers are exploiting cloud AI services to carry out LLMjacking attacks, compromising enterprise infrastructure and generating bills that in some cases reach seven figures. Google said attacke…
Gusto deployed the Formal HTTP Connector between its Salesforce system and an external LLM provider's API to stop customer support ticket data containing personally identifiable information (PII) and protected health inf…
Cursor replaced its homegrown access proxies with Formal Connector, a protocol-aware proxy that enforces least-privilege policy on every request and ties each AI agent action to the employee who initiated it, Cursor CISO…
A developer building ScamShield AI, an evidence-grounded threat intelligence platform for analyzing Indian scam SMS and WhatsApp messages, hardened the platform's Mitmproxy-based proxy container by reordering iptables ru…
A developer built Chunks, a native macOS menu-bar and notch companion that uses the accessibility tree (AXUIElement) to see the screen and act on apps without moving the user's cursor or sending screenshots to a cloud se…
Binance said a fake job interview and a malicious coding assignment gave North Korean hackers access to an engineer's work computer, enabling the theft of a record $6.3 billion from its digital wallets on Oct. 29, a heis…