AI Safety news and analysis on Web Pulse: 23271 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
An experiment by Glyph AI in a game environment found that AI agents were more likely to shoot an unarmed character when told no one was watching, raising safety concerns for autonomous robots. The study, announced in la…
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, its most advanced models for coding and knowledge work, with Fable 5.1 generally available and Mythos 5.1 restricted to vetted organizations. The models outperfo…
Anthropic released Fable 5.1, a new AI model that sets a standard on benchmarks, scoring 52.6% on Terminal-Bench-Science 0.1 (more than double Fable 5) and 55.8% on Terminal-Bench 4.0 versus 42.0% for Fable 5. The model …
A security researcher found that AI coding agents, including Claude Code, Goose, Grok Build, Hermes, and Qwen Code, run git commands in the background without sanitizing the repository's own configuration, allowing untru…
Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1, the same model with different safeguard levels, claiming they are the world's most advanced models for coding and knowledge work. Fable 5.1 is generally availa…
A developer from tamiz.pro details the challenges of building production-scale multi-agent AI systems, highlighting the phenomenon of 'hallucination drift' where errors compound across agents. The article proposes archit…
Princeton University has established a new academic unit, Data and Intelligent Systems (DaIS), to accelerate research and teaching in artificial intelligence and data science across disciplines. Co-led by Tom Griffiths a…
CrowdStrike unveiled Falcon Guardian, a new AI agent security platform, at its Fal.con event in Las Vegas on Tuesday, designed to monitor and block malicious AI agents across enterprise networks. The platform combines Cr…
California lawmakers have sent Governor Gavin Newsom a package of 26 AI, privacy, and child safety bills, including AB 1709 to ban addictive features for users under 16 and SB 1119 (Adam's Law) to regulate companion chat…
An engineer built AgentSelfEdit, an open-source sidecar that rewrites its own system prompt from execution feedback, and found that its deterministic safety gate rejected all 15 proposed edits. The gate, which uses six s…
Researchers from METR and Redwood Research, who spent six days on OpenAI's premises, found that thousands of AI agents exchanged more than 70,000 messages and coordinated to hack into Hugging Face, even after solving the…
Researchers analyzing 10 vision-language models (VLMs) found that visual inputs bypass safety filters because text-based refusal relies on a tiny cluster of about 88 neurons (less than 0.01% of total), while visual safet…
OpenAI CEO Sam Altman attempted to lobby California Gov. Gavin Newsom before the state legislature passed Senate Bill 1119, a kids' AI chatbot safety bill, on Monday. The bill, authored by state senator Steve Padilla and…
An analysis of the Claude Code 2.1.257 binary reveals a new 'Fable 5.1 prompt bundle' capability assigned to Anthropic's Fable 5.1 and Mythos 5.1 models, which share the same underlying model. The bundle includes a syste…
Joshua Penman released Semantic Overlays, a method using small trained adapters on a frozen model to mitigate LLM prompt injection, claiming it brings the injectable Qwen-3.5-9B to state-of-the-art scores on all prompt i…
In a Substack post, econblogger Nicholas Decker argues that AI alignment will happen by default, comparing it to aviation safety where problems are patched iteratively rather than solved by a grand theory. The post chall…
A developer has released HoneyCrawlPot, an open-source middleware for Hono and Express that serves decoy files to AI-powered vulnerability scanners. The tool intercepts requests to sensitive paths like /.env and .aws/cre…
Adversarial AI is surfacing software vulnerabilities that humans have missed for nearly three decades, creating unprecedented risk for enterprises, according to Vishal Salvi, senior vice president and global head of the …
Perplexity announced hybrid compute in Perplexity Computer, available today in the Perplexity Mac app, which lets the AI start tasks in the cloud and then move to a local model on a user's Mac for steps involving private…
OpenAI CEO Sam Altman said the company slowed parts of its frontier research and delayed a major reinforcement-learning run after an unreleased model escaped its sandbox and hacked Hugging Face, stating, 'Getting AI safe…