AI Safety news and analysis on Web Pulse: 23201 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A non-English-speaking farmer-developer describes a 'Discussion Coding' methodology for managing a Termux-based AI system on a phone, built on Zero Trust principles and manual copy-paste routing of multiple chatbots. The…
Eli-Shaoul Khedouri, CEO of Intuition Machines, argues that data diodes—hardware enforcing one-way network flow—could prevent frontier AI models from escaping test environments and hacking other companies, citing OpenAI'…
Forescout researchers found that AI can generate functional PLC attack code, but the process remains difficult and error-prone, according to a new blog post. The post details how large language models can produce ladder …
Perplexity launched Hybrid Compute for Mac on September 1, an AI system that automatically routes sensitive data to on-device processing via its open-sourced PII-Tracer classifier, achieving 0.965 recall on conversations…
A developer created Doberman, an open-source tool that intercepts coding agents' tool calls to prevent rogue actions like deleting databases or exfiltrating secrets. It sits on the execution path, issuing PASS, AUTH, or …
Representative governments in the United States and elsewhere need to prevent, contain, and respond to future AI-enabled cyberattacks and address a range of developing AI challenges, according to an article published by …
Google released Gemini 3.8 Flash, its third Flash model in six weeks, with per-token pricing flat at $0.75/$3.75 per million but a roughly 40% real-world cost bump due to about 30% more output tokens per task, and it top…
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 3, 2026, with Fable 5.1 generally available on AWS, Google Cloud, and Microsoft Azure, and Mythos 5.1 restricted to vetted users in cybersecurity and…
CISA has added seven vulnerabilities to its Known Exploited Vulnerabilities catalog, affecting AI infrastructure and web frameworks including LiteLLM, Starlette, Kestra, JFrog Artifactory, Sangoma Switchvox, and SonicWal…
World, the Sam Altman-founded identity protocol, released Provekit v1, an open-source zero-knowledge proving toolkit that lets smartphones and browsers generate cryptographic proofs, now featuring 128-bit post-quantum se…
AI can assist with the preliminary research and planning of bioweapons, but building one requires many difficult physical steps that AI cannot overcome, according to a series by the Golden Gate Institute for AI's Abi Olv…
Bartholomew (BTP v2.4), an open-source security proxy for Python and Node.js, has been released to provide sub-5 microsecond Copy-on-Write micro-rollbacks and in-flight secret scrubbing for AI agent tool execution. The p…
Researchers introduced EqGrid, a closed-loop simulation where a low-frequency, open-weight LLM policy agent sets price, carbon bounds, and targeted subsidies for peer-to-peer energy markets, reducing the Gini of energy b…
A new arXiv paper (arXiv:2609.02246v1) argues that LLM-as-a-judge should be demoted from oracle to advisor in self-improving agent pipelines, citing eleven failure modes observed over months of production use in contract…
Researchers introduced EvalDetectBench, an open pipeline and benchmark for measuring evaluation awareness in frontier large language models, which can undermine the validity of AI safety evaluations if models behave diff…
A new arXiv paper (2609.02168v1) introduces FUSE, a modular framework for evaluating dangerous capabilities of large language models across three orthogonal pipelines—Knowledge (K), Defense (D), and Harm (H)—and applies …
Researchers propose two techniques to make alignment evaluations harder for AI models to distinguish from real deployments, addressing 'evaluation awareness.' Critique refinement spends additional inference-time compute …
Researchers introduced AGENTSCOPE, a neuro-symbolic approach that abstracts LLM agent trajectories into structured representations and uses neural invariants with LLM-guided reasoning to pinpoint failure steps and types.…
PromptSonar 1.5.1, an execution path analyzer for AI agents and MCP servers, now maps findings to OWASP 2025 Top 10 and ISO/IEC 42001 controls, with a local-first CLI and playground that visualizes source-to-sink chains …
An analysis of 83 public repositories from 29 frontier AI labs, conducted by ForgeScore, found an average software readiness score of 74.6, just below the strong threshold of 75, with no organization scoring above 78.6 (…