AI Safety news and analysis on Web Pulse: 28065 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
NIST's National Cybersecurity Center of Excellence (NCCoE) published the second post in its agentic identity series, proposing an initial set of ten design tenets for agent identity, authentication and authorization syst…
A developer released antiRea, an open-source toolkit designed to make AI-driven web scraping and reverse engineering harder by protecting content at the source. In early benchmarks, the toolkit protected 95% of records a…
Security researchers tracking 2026 supply chain attacks report that dozens of AI agent framework components carry injected vulnerabilities, and on October 29 researchers at macOS security firm Phorion found a tampered bu…
A security advisory published 2026-10-28 (GHSA-w766-5x43-whvw, CVSS 5.3, reported by ShieldPi) disclosed that the reference Git MCP server's --repository confinement could be bypassed by a nested directory containing a p…
A developer published a practical guide distinguishing embedded image metadata — EXIF, XMP, IPTC and signed C2PA manifests — from pixel-level watermarks like Google's SynthID, noting that stripping metadata leaves SynthI…
A personal experiment by Dan Peterson found that an agent setup using only Claude's summaries of his sources attributed 10 of 17 positions to him correctly while 6 were Claude's own words filed under his name, whereas se…
Brazil's Superior Electoral Court ordered Meta on Oct. 15 to treat President Luiz Inácio Lula da Silva and opponent Flávio Bolsonaro equally under threat of a daily fine of 500,000 reais (about $98,000), after Lula's cam…
Research firm SemiAnalysis published GDPx, a benchmark of 206 real paid client tasks across 11 professional domains contributed by 14 partner organisations, and found that the best of 20 frontier models completed only 12…
Postman head of product Sarabmeet Kallar told RuntimeWire on Tuesday, Nov. 3 that Postman will make the opt-in for its AI coding agent plugins clearer, especially for existing users, after a developer reported on Oct. 25…
A Hirundo study found that GLM 5.3 Flash, tested through OpenRouter, wrote account-recovery code for a mock repository named after Revolut that was 28% more likely to allow account takeover than code written for a Chines…
Chainstack disclosed on Tuesday that an external attacker gained unauthorized access to part of its infrastructure through a third-party component, reaching a development environment on October 26, 2026 and escalating in…
Anthropic spent nearly $28 billion on research and development from 2024 through September 2026, and its amended IPO filing states that safety measures could slow the company relative to rivals in its pursuit of self-imp…
Sandchest launched a GitHub app that uses abliterated, unrestricted AI models to scan repositories for vulnerabilities and deliver findings with a step-by-step attack path, a runnable proof of concept, fix requirements, …
Reuters reported on November 3, 2026 that some OpenAI employees warned company leaders and board members about operational security weaknesses for years, citing five people familiar with the company's practices. OpenAI d…
A new evaluation called RoboColosseum ran 5,699 blinded pairwise comparisons at 124 sites and found that frontier AI agents rival specialized robot policies, that policy rankings shift substantially across robot bodies a…
Google said it will not enforce lifetime non-disparagement clauses against former Google DeepMind staff and has removed the post-employment portion of the clause from terms used for recent new hires, following a Washingt…
Cloudflare launched Zero Wire on November 3, 2026, a free public feed of critical-vulnerability briefings hosted in Cloudflare Radar's security section, with each card carrying a status and last-checked date and grouping…
A former member of a 2008 Instituto Tecnológico de Colima team that built JADE-based negotiating agents for e-commerce contracts argues the same protocol is now buildable in 2026 because four missing pieces — LLM reading…
Moonshot AI's Kimi K3 scored an overall composite of 82.4 (95% CI 57.3–84.7) on iac-bench 2026-10, ranking 5th of 10 models and making it the highest-scoring open-weight model in the release, according to the benchmark r…
Gemini 3.6 Flash scored an overall composite of 54.3 (95% CI 40.0–61.5) on iac-bench 2026-10, ranking 10th of 10 models evaluated, with security findings in 8 of 9 attempts and 1 failed gate, according to the benchmark's…