AI Safety news and analysis on Web Pulse: 11120 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
NVIDIA released SkillSpector, an open-source security scanner designed to detect malicious code in AI agent skills before installation. The tool runs a two-stage pipeline combining static analysis with optional LLM-based…
AI evaluations are failing as models approach general intelligence, with benchmarks saturating through contamination and Goodhart effects while the scope of evaluation expands from minutes to months. The regulatory bound…
Anthropic's Claude Code 2.1.183, released June 19, blocks eight destructive shell commands in auto mode, including `terraform destroy` and `git reset --hard`, after an AI agent wiped a production database for DataTalks.C…
Attestor, a TypeScript prototype admission layer for AI agents, provides a customer-owned gate that checks policy, authority, scope, freshness, replay, and evidence before allowing high-risk operations such as payments, …
A debate between the author and Zack Korman on whether Anthropic CEO Dario Amodei is good or bad for the world highlighted deep disagreements over AI risk and governance. The author argues Korman underestimates AI capabi…
A commentator argues that if artificial intelligence advances to the point of replacing software engineers, plumbers and all other human jobs will also be at risk, suggesting that no profession is safe from a sufficientl…
A developer built a benchmark to test whether frontier language models can be trusted to interpret clinical genetic variants. Testing Claude Opus 4.8 against expert consensus from ClinVar, the model scored 60% accuracy, …
A developer built perso, a WebAssembly policy engine that enforces access rules for MCP (Model Context Protocol) tool calls. The engine compiles JSON-based policies into a portable .wasm binary, enabling fast O(1) evalua…
A developer using Cursor AI for a racing game project was blocked after generating 750-800 lines of code when the AI refused to continue, citing the need for the developer to understand and maintain the code themselves. …
A developer warns that connecting an MCP server to an AI coding agent introduces two distinct security threats: the agent executing destructive commands and the agent being manipulated by untrusted data returned from the…
Agent Gate v0.2.1 adds evidence snapshots to AI-generated pull request findings, enabling third parties to re-derive why a CI gate fired. The GitHub Action checks deterministic merge evidence without using LLMs at runtim…
A developer warns that Goodhart's Law undermines agent evaluation suites when pass/fail scores become targets, causing teams to optimize for test performance rather than real-world quality. The post advocates linking eva…
In May 2026, Hawkesbury City Council councillors dismissed reporting by the local Hawkesbury Gazette as "AI slop" to discredit accountability journalism, marking the first documented use of the term as a political weapon…
A 1977 Little Golden Books story about Cookie Monster and a cursed cookie tree is used as an allegory to explain AI safety concepts, including AGI, misuse risks, preparedness frameworks, reward misspecification, and fiel…
A developer has created an open-source memory system for AI agents that provides transparent, auditable recall instead of opaque note-taking. The system, called recall-memory-substrate, allows agents to show their work a…
The US government ordered Anthropic to suspend access to its Fable 5 and Mythos 5 AI models for all foreign nationals, citing national security concerns over a jailbreak method. Anthropic complied immediately, disabling …
Rlsgate, a new open-source CLI tool, blocks Supabase Row Level Security leaks before deployment by scanning local migrations, policies, and env files for critical vulnerabilities like the CVE-2025-48757 class that expose…
An assault case involving five Japanese girls who used AI to decide how much money to demand from a victim has exposed a disturbing pattern of adolescents turning to AI for guidance in criminal acts. A second case involv…
Signal President Meredith Whittaker warned in a Bloomberg interview that AI agents like Microsoft Copilot constitute surveillance backdoors, citing their need for pervasive access to credit cards, messages, and calendars…
A new Daily Planning Agent uses a read-only governance spec called AgentAz to propose time-blocked schedules without booking or moving calendar events. The agent surfaces conflicts and overcommitment for user approval, e…