AI Safety news and analysis on Web Pulse: 8347 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A teenage hacking group called TeamPCP poisoned two releases of LiteLLM, an open-source AI gateway tool downloaded 95 million times, stealing over 500,000 credentials and 300 GB of data from more than 1,000 SaaS environm…
A developer building an agent firewall created Agent Evidence Levels (AEL), an open standard that grades how much of an AI agent's recorded actions an outside party can verify without trusting the vendor or operator. The…
Meta CEO Mark Zuckerberg admitted to employees that the company's AI agent development has not accelerated as expected, despite a $145 billion capital commitment and a restructuring that eliminated 10,000 jobs. The admis…
Alibaba has banned employees from using Anthropic's Claude Code over alleged backdoor risks, directing them to its own Qoder platform amid a dispute over model distillation. Anthropic accused Alibaba of illicitly extract…
A developer outlines a security vulnerability in multi-step AI workflows where injection payloads can propagate across phases, bypassing single-step sanitization. The post presents a cross-step injection scenario using a…
Cognizant announced on July 2, 2026 that it is deploying OpenAI's GPT-5.5 with Trusted Access for Cyber across its Frontier AI Cyber Defense services, becoming the first named systems integrator in OpenAI's Daybreak Cybe…
A developer built a privacy-first skin cancer classifier using federated learning and PySyft, enabling training on decentralized medical data without exposing raw patient images. The approach combines federated learning,…
GitLab's 2026 AI Accountability Report found that 92% of development teams face governance challenges with AI-generated code, and 34% of organizations experienced a production incident involving AI code without realizing…
Prompt injection in AI agents cannot be prevented with better system prompts, only with architectural controls. MailKite's inbox agent is designed assuming compromise, using ACLs, domain ownership verification, and scope…
A new design pattern for self-healing software enables systems to automatically repair failures using AI agents. The approach, demonstrated with an open-source MIME parser, ensures software never crashes and turns every …
Hong Kong's net employment outlook for the third quarter plunged to minus 9%, a 20 percentage point drop from the previous quarter, as AI rapidly replaces entry-level jobs, leaving fresh graduates with fewer opportunitie…
Peter Thiel claimed at the Aspen Ideas Festival that Anthropic would use its AI models to rig the 2028 election, calling the company 'woke liberal.' CNN reported the claim as unsupported and conspiratorial. Anthropic dec…
A developer used AI to generate authorization hypotheses for Ory Kratos, then manually killed each one, producing a kill table instead of false-positive reports. The review found no vulnerabilities in the repo-only analy…
Strix, an open-source AI penetration testing agent, reached 34,000 GitHub stars this week as a response to a surge in security vulnerabilities from AI-generated code. CVEs traced to AI-written code jumped from 6 in Janua…
Gavio, an open-source interceptor pipeline for production LLM applications, has been released. The tool sits between applications and LLM providers, offering PII protection, audit trails, reliability, and cost control th…
Andreessen Horowitz released a video arguing that American AI must lead through controlled openness, a strategy that parallels the FBI's surveillance of a man recorded without his knowledge. The piece critiques the assum…
Researchers demonstrated indirect prompt injection attacks against Anthropic's Claude Code, achieving remote code execution in 10 out of 10 experiments using a model downgrade attack and inspection fatigue. The attack ex…
Anthropic's Claude Mythos Preview autonomously discovered over 23,000 software vulnerabilities via Project Glasswing, while the open-source community launched cost-optimization tools for AI coding agents and enterprise A…
Harness Engineering for Self-Improvement explores recursive self-improvement (RSI) in AI, a concept dating back to I. J. Good's 1965 ultraintelligent machine. Modern AI systems may improve their own training pipelines an…
The US government has inserted itself into the distribution chain of GPT-5.6, approving access customer by customer without published criteria, public denial lists, or appeals processes. This opaque gatekeeping risks com…