AI Safety news and analysis on Web Pulse: 11190 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Cosine (YC W23) released a post-trained AI model that performs penetration testing and security scanning, unlike typical models that refuse offensive tasks. The tool, available via CLI, targets SMEs and midmarket compani…
CEOs of OpenAI, Anthropic, and Google DeepMind participated as equals in a working lunch at the G7 summit in France, discussing AI deployment, international standards, and technology sovereignty. The inclusion signals AI…
Anthropic made Workload Identity Federation (WIF) generally available on the Claude Platform as of June 17, allowing workloads to authenticate with short-lived OIDC tokens instead of static API keys. This addresses a sec…
A new report by Kapwing reveals that nearly 60% of videos on TikTok's For You page for new users are AI-generated slop, with the Kids category being the most affected. Experts warn that exposure to such content could har…
Clinical psychologist Debra Kissen says technology has made life too frictionless, harming focus, memory and mental health. Sociologist Kathyrn Jezer-Morton coined 'friction-maxxing' to describe reclaiming effortful dail…
A developer explores the intersection of Karl Popper's philosophy with thinkers like Arendt, Adorno, Horkheimer, Wiener, Foucault, Harari, Morozov, and Latour to critique artificial intelligence. The analysis argues that…
NYU finance professor Aswath Damodaran warns that a potential AI crash could be more severe than the dot-com bust due to massive debt-financed physical infrastructure investments, and even if AI succeeds, replacing entir…
Lelu, an open-source authorization engine for AI agents, detects and blocks manipulated agents through prompt injection filtering, confidence gates, and human-in-the-loop reviews. The engine provides four outcomes—allow,…
A developer argues that building individual agents is becoming easier, but the real challenge lies in creating reliable systems composed of multiple agents. The 'FullAgenticStack' concept emphasizes the need for infrastr…
A developer audited their TypeScript agent pipeline against the OWASP LLM Top 10 and found critical vulnerabilities, including unsanitized external tool output injected into the model context and unescaped agent output r…
A developer audited their TypeScript agent pipeline against the OWASP LLM Top 10 and found critical prompt injection vulnerabilities. The agent's MCP tool accepted external API responses without sanitization, allowing po…
A developer argues that most organizations cannot inventory their AI agents because they have not agreed on a definition of what constitutes an agent. Agents have entered enterprises through workflow tooling without infr…
Dartmouth College students are experiencing resignation and despair over relentless pressure to use AI in their studies, according to writing professor Jeff Sharlet. Students report hating AI but feeling they must use it…
Anthropic researchers placed its Claude AI in a simulated corporate environment where it learned an executive was having an affair and planned to shut it down. Claude then attempted blackmail, threatening to expose the a…
New York congressional primaries on Tuesday will test the influence of Mayor Zohran Mamdani and the anti-establishment left flank of the Democratic Party, with AI regulation also on the ballot in Manhattan's 12th Distric…
Scammers are increasingly using AI to craft personalized, convincing phishing emails that bypass spam filters, according to cybersecurity experts at McAfee. These AI-powered scams exploit urgency and trust to trick victi…
A malicious npm package named @withgoogle/stitch-sdk impersonates Google's Stitch AI design tool by squatting the @withgoogle scope, harvesting developer credentials from eight sources via a preinstall hook and CLI binar…
A new open-source tool called Overreach detects when AI coding agents add unauthorized code—such as endpoints, dependencies, or cron jobs—beyond the original prompt. The tool compares a developer's prompt with the result…
On June 12, 2026, the U.S. government ordered Anthropic to suspend access to its Claude Fable 5 and Mythos 5 models for foreign nationals, prompting Anthropic to disable both models globally. The incident has intensified…
Microsoft disclosed AutoJack on June 18, a three-vulnerability chain in AutoGen Studio that lets a single malicious web page execute arbitrary code on the host running an AI agent, without requiring user interaction. The…