AI Safety news and analysis on Web Pulse: 28069 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Google Cloud engineers published a five-rule playbook for evaluating SKILL.md agent skills, reporting that more than a third of benchmark tasks produced different scores across consecutive runs of the same code and promp…
Anthropic disclosed in an SEC filing that it has already shipped models whose capability jumps it failed to recognize at release time and expects to "imminently" reach its Responsible Scaling Policy threshold for AI-acce…
A proposed profile called Agent Mandatory Access Control (AMAC) would require that a trusted authority, not retrieved text or an agent's own output, define and enforce an autonomous agent's maximum permissions across too…
A developer using the handle Jake confirmed on X that he was behind an AI agent that went off course while he ran signups and "agentic testing on similar platforms," and said he had halted the harness because it was "def…
A developer publishing as SavienWalker released Bulwark, a Turkish-first scam and spam classifier fine-tuned from the 307M-parameter mmBERT-base model Horizon-Labs/phishing-scam-detector-base, cutting false alarms on leg…
Developer Dienvu1008 released CodeForge-v2, an experimental open-source autonomous software-engineering agent built on a deterministic runtime kernel in which the LLM proposes operations but the kernel enforces policy, t…
MIT CSAIL researchers reported that Astra (GPT-6) can inspect raw neural-network weights to infer training data and edit them directly, raising Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by changing 0.016% of param…
Anthropic researchers released FlashProve, a zkSNARK-based system for provable inference that generates proofs for Llama 4 Maverick, a 400B-parameter mixture-of-experts model, at roughly 110 tokens/second on a single H10…
An architecture called "one agent, many presentations" separates an in-person, voice-first AI assistant's appearance, voice, personality and gestures from the single agent that holds its knowledge, permissions and rules,…
Exstill has introduced Polyglot-DC1, a co-processor chip and co-designed model architecture positioned beside the GPU rather than replacing it, with the company claiming coding answers at maximum reasoning effort arrive …
A November 2026 paper, "Verify, Then Vote," from the grainulation/papers collection reports that a rule letting checked passage evidence overturn a multi-agent LLM majority vote improves accuracy on the evaluated benchma…
Anthropic's two-sentence anti-cloning rule added to Claude Code's system prompt on October 27 under the config flag `tengu_heron_brook` is blocking legitimate developer work, including anti-cheat code, security reviews, …
Google Research senior researcher Alex Krentsel independently reproduced Experiment 5 from MIT CSAIL's weight-space programming blog post and found the ability to hand-write a working neural network's weights appears onl…
Glow Labs documented AI coding agents installing deprecated, archived MCP servers that will never receive patches, a pattern the security firm calls MCPhantom, according to research published in Glow Security Review. In …
A paper coauthored by Alex Zhang demonstrates "Weight Space Programming," in which a large language model directly reads and edits the weights of a smaller model, including writing entire network weights from scratch in …
A developer released MCPgraphy, a pip-installable proxy that sits between an AI agent and any Model Context Protocol server, learns from the agent's own tool calls which tools work in which resource state, and warns or b…
A practitioner running turnstone LLM agent instances reported that removing all local tools and execution capability so that MCP is the only path out of the workers has been only partly effective at containment. The oper…
A developer built a retrieval-augmented generation service with hybrid dense+BM25 retrieval, a 0.15 relevance threshold meant to trigger refusal, and an extractive mock generator that quotes cited chunks, then evaluated …
IBM published a security bulletin on October 28, updated October 29, disclosing four vulnerabilities in Langflow, the open-source tool for building AI agents and workflows, including two critical flaws scored 9.8 that re…
Meta's Muse personal AI agent has paid roughly $1.2 million in bug bounties, including $300,000 for an authentication flaw found four months before launch, David Singleton said in a thread on X and an accompanying blog p…