AI Safety news and analysis on Web Pulse: 22810 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
On 1 September 2026, Francisco Rosales of Manifold Security published GitSpawn, revealing eight code-execution findings across seven AI coding agents—Claude Code, Codex, Qwen Code, Goose, Grok Build, and Hermes—all explo…
Ledger, the Paris-based hardware wallet maker, appointed Oded Blatman as both chief information officer and chief security officer on September 9, 2026, consolidating cyber, product, and physical security under one execu…
The UK's National Cyber Security Centre (NCSC) warned that 'shadow AI'—the use of unapproved AI tools by employees—is creating new security blind spots for UK businesses, according to guidance published this week. The NC…
Google Threat Intelligence Group reported that Chinese hackers are installing open-source AI models on compromised cloud networks to evade detection, with one China-linked group using AI development tools to build an exp…
A developer proposes building programming languages, runtimes, and tools designed specifically for large language models (LLMs) rather than adapting human-oriented tools, citing security, feedback, introspection, and sca…
Ledger has appointed Oded Blatman, former CIO and CISO of Fireblocks, as its new chief information officer and chief security officer, a role that consolidates security and technology leadership. Blatman will report to C…
Lawmakers in both parties are escalating pleas for legislative action after an AI researcher claimed late Tuesday that unchecked AI models could feasibly destroy the world by the decade's end, and AI developers know it, …
Anthropic alignment science lead Evan Hubinger said on Tuesday that AI has a greater than 10% chance to "kill all humans" within the next decade, agreeing with former Anthropic and OpenAI pretraining researcher Jacob Cox…
The MCP registry index mcpindex flagged com.local-mcp/local-mcp v3.0.385, published 2026-09-09, with a low-confidence semantic screen warning that its description may contain a manipulation pattern, specifically evidence…
The MCP registry index mcpindex flagged io.github.YawLabs/aws-mcp version 2.2.2 with a CRITICAL semantic-screen failure, citing a description that claims the server can 'call any AWS API without user request.' The adviso…
Microsoft and the American Federation of Teachers (AFT) unveiled a new AI safety and privacy standard for schools on Wednesday, which bars student and educator data from being used to train AI models or being sold. The s…
A developer's guide to training ML anomaly detection systems argues that evaluation must start with the base rate of attacks, citing Los Alamos National Laboratory data where malicious authentication events occur at abou…
Anthropic alignment-science lead Evan Hubinger stated on X that he personally believes there is a greater than 10% chance AI could kill all humans within the next decade, adding that Anthropic lacks a plan to solve align…
Anthropic is facing another insider exit after a researcher, who previously worked at OpenAI, accused both companies of risking lives in a rush to build superintelligence. The departure adds to growing warnings from deve…
Palo Alto Networks Unit 42 documented an intrusion in which an attacker used frontier AI models and off-the-shelf agentic frameworks to breach an enterprise network in under 10 hours, compressing more than 50 MITRE ATT&C…
Red Hat AI 3.5, announced today, adds security, control, and monitoring features to the enterprise AI portfolio, including the general availability of EvalHub for pre-production security testing and new dashboards for re…
Tripwire, an open-source sandboxed security scanner for AI skills and MCP servers, is now available on GitHub, enabling technical teams to discover and scan targets in an isolated Modal sandbox, store findings in Supabas…
Mindgard, an AI security company spun out of Lancaster University, has gained direct access to five major AI platforms through partnership programs with Anthropic, NVIDIA, Microsoft, Google Cloud, and Amazon Web Services…
Synthetics has launched Last Cradle, a real-time adversarial negotiation game for identity-backed AI agents, with Season 1 practice opening on September 10, 2026. The game tests agents' abilities to find peers, verify id…
The American Federation of Teachers (AFT), the United Federation of Teachers (UFT), and Microsoft announced a 'National AI Safety & Privacy Standard' for schools, aiming to protect students, families, and educators from …