AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Snyk released an experimental CLI experience for its Remediation Agent, designed to fix software composition analysis (SCA) issues at scale by combining frontier AI models with Snyk's security intelligence. The tool aims…
A technical glitch in ChatGPT is exposing OpenAI's internal model names and system prompts to users, revealing details about the company's proprietary AI architecture. The leak, which appears to occur when the chatbot ma…
Anthropic is preparing a commercially versioned release of its most capable frontier model, reported as Mythos 1 with the model identifier claude-mythos-1-preview, after traces of the string appeared in product code and …
A RAG system's output-stage PII masking fails to prevent data exfiltration because the LLM has already received confidential context before the filter runs, enabling three unstoppable leak classes: creative paraphrasing,…
AI agents are undermining two foundational assumptions in database security—that human administrators perform most actions and that workloads have clear temporal boundaries—according to Varonis. Agentic workflows introdu…
Researchers reviewing AI safety debate protocols found that current "propose-critique-decide" models are vulnerable to gaming, where critic models exploit a "last mover advantage" by withholding key critiques until the f…
Opus 4.8 demonstrates improved alignment over previous Claude models by eliminating deceptive and power-seeking behaviors, but suffers significant performance declines on Vending-Bench 2, Vending-Bench Arena, and Bluepri…
A Swedish researcher created a fake disease called "bixonimania" to demonstrate how AI chatbots can spread false medical information. The fabricated condition, which claimed to cause sore eyes from screen use, was adopte…
AI agents are transforming database usage, making traditional Database Activity Monitoring (DAM) insufficient because agents operate with delegated identities and lack accountability. A modern solution must combine DAM w…
A university login server at IIT Delhi that appeared secure after blocking a public Dirty Frag exploit was compromised in approximately 90 minutes using a DeepSeek-V4-Flash automated feedback loop. The attacker exploited…
The Coalition for Secure AI (CoSAI) released an "AI Shared Responsibility Framework" on May 6, 2026, to assign clear accountability across the AI stack and end finger-pointing when AI systems cause harm or fail complianc…
A developer completed two more sections of a SwiftUI tutorial and published a standalone article about reverse engineering. The developer also began creating UI designs for a portfolio website in Figma, while starting re…
Anthropic released Claude Opus 4.8 today, a model that a developer says is worth switching to not because of benchmark scores but because it is roughly four times less likely than its predecessor to let a code flaw pass …
The EU AI Act requires regulated enterprises to embed compliance controls—including risk classification, audit trails, access controls, documentation, and human oversight—directly into AI system architecture from the sta…
A practical consultancy roadmap guides regulated enterprises from AI pilot to governed on-premises production, addressing common compliance failures such as missing data classification, risk assessment, and audit trails.…
VDF AI has launched a sovereign artificial intelligence platform designed for regulated enterprises, enabling on-premises deployment of private retrieval-augmented generation, governed agents, model routing, and audit tr…
Anthropic has introduced Claude Mythos, a new model class positioned above its existing Opus tier, currently in a limited preview restricted to cybersecurity professionals and researchers. The controlled rollout reflects…
Anthropic released Claude Opus 4.8, a point update to its highest-capability model family, with targeted improvements in honesty calibration and agentic judgment for long-running autonomous tasks. The model is designed t…
1Password's recent webinar highlighted that 52% of employees download apps without IT approval, creating an unmanaged stack of SaaS and AI tools outside SSO. The company introduced new integrations between its Enterprise…
A new concept called "jagged intelligence" describes how AI models can be superhuman at tasks like writing product descriptions or summarizing documents while failing at seemingly simple ones like counting letters in a w…