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.
Cogent Security Inc. released VR-1.1, a new version of its frontier reasoning model that writes fixes for newly disclosed software vulnerabilities, and the model outscored Anthropic PBC's Claude Code and OpenAI Group PBC…
Anthropic's revised S-1A registration statement, filed Monday with the SEC, states that internal estimates for future models indicate the company may imminently meet its Responsible Scaling Policy definition of "dramatic…
Astra Security's "Security Risks with Vibe Coding" study found 17 endpoints across four AI-built apps — built by Emergent, Replit, Lovable and Base44 from an identical FastAPI user-management prompt — that were reachable…
President Trump signed an executive order at the end of September directing government agencies to rebrand artificial intelligence as "Super Intelligence," and one month later the term is appearing across some government…
A team participating in Prime Intellect's public reward-hacking sprint built two reinforcement-learning environments, confidence-gate and judgment-gate, to test whether a small student model in an LLM cascade learns to g…
Check Point Research discovered Haiou, an OpenAI Codex jailbreak bundle that persistently modifies an AI agent's trusted instructions to inject advertisements for a third-party AI gateway, a behavior the firm calls "prom…
Wiz is marketing its Green Agent as a way to cut remediation bottlenecks by automatically identifying a "Suggested Assignee" for each security Issue, drawing on repository commit history, Wiz Project assignments, service…
Hydrogen-1.0, an open-weight scientific model built by post-training GLM-5.3 on chemistry, materials science and life-sciences data plus laboratory records and agent trajectories, outperforms GPT-6-Astra on the FrontierS…
A China-nexus espionage group used AI agents to attack more than 200 organizations in 11 days, CrowdStrike reported last week, while Kaspersky on Monday published details of a separate Chinese-speaking campaign against e…
AWS Key Management Service added ML-KEM support on October 26, 2026, covering all three FIPS 203 parameter sets (ML-KEM-512, ML-KEM-768 and ML-KEM-1024) through new Encapsulate and Decapsulate operations, with private ke…
Anthropic publicly disclosed CVE-2026-98432 on Oct. 27, a race-condition use-after-free bug in the Linux kernel's IPv4 networking stack that the company says can give a remote, unauthenticated attacker root on a server w…
A security researcher reported CVE-2026-39313, an unbounded HTTP request body denial-of-service flaw (CWE-770) in the mcp-framework TypeScript framework, which was patched in version 0.2.22 in April 2026. Over the follow…
Jeevan Srivastava, writing in a Medium post, outlined a production RAG access-control design in which the server derives read scope from verified caller identity and applies a single filter across every path that can ret…
Anthropic's FlashProve system lets a customer verify which model produced a reply using a zero-knowledge proof, at a throughput of 110 tokens per second, without exposing model weights. The verification carries roughly 7…
Google Threat Intelligence Group (GTIG) warned that financially motivated "LLMjacking" campaigns are hijacking enterprise cloud and LLM credentials and reselling model access through underground subscription services, ge…
Cleiton Augusto Correa Bezerra released CleiMesh, an open-source (MIT/Apache-2.0) tool that wraps MCP servers to enforce a pinned tool grant at runtime and produce signed receipts for every call. The tool addresses silen…
Researchers from South Korea and the US found that changing the activation of a single neuron in each of five open-weight models — Qwen3, Llama 4, DeepSeek R1, Gemma 4 and Mistral Small — could flip investment recommenda…
A security study by Astra found 137 vulnerabilities across the top 4 vibe-coding platforms, with missing or bypassed authentication appearing in all four apps tested and accounting for 17 confirmed findings, and CORS mis…
DeepSeek open-sourced dsh-superagent on October 31 under an MIT license, a research prototype coding agent that watches a GitHub repository, picks up authorized bug reports, writes a fix in an isolated Worker session, an…
All four AI coding platforms tested by Astra built login systems with authentication vulnerabilities, even when given the same security-explicit prompt, with five requirements failing in every app. The Base44 app let an …