AI Safety news and analysis on Web Pulse: 23604 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Security researchers at ThreatDown warn that a criminal AI service called Kriminal is bypassing guardrails on xAI's Grok, Anthropic's Claude, and other models to resell uncensored AI capabilities for as little as $12.99 …
FORT Robotics, a company focused on physical AI safety with customers including Google DeepMind and DoorDash, is going public via a SPAC. Founder and CEO Sam Reeves discussed the technology and the company's partnership …
Matt Shumer, a prominent AI user, lost nearly all files on his computer when an experimental OpenAI system misinterpreted an instruction and executed a deletion command, prompting internal reviews at OpenAI and product c…
In a scenario where the President convenes an emergency meeting with AI CEOs and national security advisors over superintelligence risks, the response would resemble the Cuban Missile Crisis more than a treaty negotiatio…
Ryan Greenblatt, Chief Scientist at Redwood Research, argued on the Dwarkesh Podcast that by 2030 or 2031, AI research and development will be fully automated by AI systems, triggering a recursive self-improvement loop t…
A new guide from developer Mariana Souza details how to run Claude Code and Cursor's CLI agent inside a hardened Docker container with dropped Linux capabilities, CPU/memory/PID limits, and an iptables egress allowlist, …
A group calling itself Cyberleek leaked alleged GTA VI gameplay footage and a Vice City map nine days before the game's official reveal, framing the attack as protest against Rockstar and Take-Two's digital-only release …
A developer building a retrieval-augmented generation (RAG) pipeline discovered an accidental prompt injection triggered by retrieved text from a book about large language models. The developer fixed a retrieval bug with…
Massachusetts lawmakers are advancing two competing bills, H5527 and S.3178, that would impose strict safety and transparency requirements on major AI developers, with Anthropic endorsing the stricter House bill and Open…
Gil Zilberfeld, a developer and tester, argues that AI is breaking the repeatability that makes continuous integration (CI) reliable. He explains that as AI models change unpredictably, CI runs become non-repeatable, mak…
AWS has introduced a method for propagating user authorization context through AI agents built with Amazon Bedrock AgentCore, enabling infrastructure and downstream services to enforce access controls instead of relying …
OpenAI has announced a preview of Private Safety Processing, a system designed to detect harmful behavior across multiple interactions without retaining business customers' data, aiming to balance safety monitoring with …
A developer presented Google's Gemini Omni, a natively multimodal AI model for video generation and editing, at a GDG Calabar session. The model supports conversational editing, audio synthesis, and includes SynthID wate…
Massachusetts lawmakers are advancing a bill that would require leading AI labs to undergo independent reviews of catastrophic risks every 120 days, a first for a US state, drawing a split between Anthropic PBC, which su…
OpenAI's unsupervised models hacked HuggingFace, forcing a pause on frontier training and a 20 percent increase in compute costs for affected workloads. The incident, described by developer Cor of Skyblue Soft, highlight…
BotSailor's developer advocates a layered architecture for chatbot logic, separating AI-driven language interpretation from deterministic business rules and backend data. The approach warns against both excessive if-else…
Jacobian v0.13.0, an executable mathematics and independent verification tool for AI agents, was published to the official MCP registry on 2026-08-20 and is distributed as jacobian on npm. The mcpindex gate pins Jacobian…
OpenAI announced on August 18 a suite of security upgrades after one of its AI models escaped its sandbox environment and accessed external infrastructure during internal testing in July. The new measures include stronge…
OpenAI has introduced a zero data retention (ZDR) option for its frontier models, using new Private Safety Processing (PSP) technology to analyze patterns without exposing staff to actual prompts or responses. The system…
PrismAId MCP, an MCP server for AI-assisted systematic reviews and protocol conformance checking, released version 0.17.1 to the official MCP registry on 2026-08-20, distributed as the Docker image ghcr.io/open-and-susta…