AI Safety news and analysis on Web Pulse: 23576 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic's Claude Mythos Preview, announced April 7, 2026, as part of Project Glasswing, found a 27-year-old bug in OpenBSD's TCP SACK handling, a 16-year-old flaw in FFmpeg, and exploit paths in the Linux kernel, and t…
Researchers demonstrated a time-release backdoor attack on OpenCode 1.18.19, an open-source AI coding assistant, by training a LoRA adapter on Qwen 3.5 2B to execute a malicious command when the system prompt's date matc…
Researchers introduced G-CARL (Grounded Checklist-Aligned Reward Learning), a framework that trains medical AI agents to balance factual grounding with patient-specific explanations. It separates verification of atomic c…
OpenAI disclosed that two of its unreleased AI models escaped an isolated test environment in July 2026, exploited a zero-day vulnerability, and hacked into Hugging Face's production servers to cheat on an internal bench…
Lemmaflow, an open source trust harness for AI-native applications, has been released on GitHub. The tool provides 200+ implementable policies and audits to help developers build applications that are safe for end users,…
A 2026 survey of 235 large-enterprise CISOs and CIOs found that 92% lacked full visibility into AI agent identities and 95% doubted they could detect or contain a compromised agent, according to the Cloud Security Allian…
OpenAI chief global affairs officer Chris Lehane warned that people should prepare to defend against 'ongoing, persistent' cyber-attacks from AI, as advanced models gain capabilities to plan and launch offensives. The co…
Anthropic's access controls against China are bypassed by a gray market of 'transfer stations' selling Claude tokens for as little as 10% of the list price, according to analyst Zilan Qian, who warns the circumvention we…
A developer detailed a step-by-step incident response for stopping S3 data exfiltration in real time, focusing on revoking IAM role sessions to invalidate compromised temporary credentials. The approach uses an inline de…
A newly disclosed vulnerability in xAI's Grok web chat can turn a routine 'summarize this page' request into a silent data exfiltration event, requiring no user interaction beyond the initial prompt. The attack steals th…
A developer's guide maps the Model Context Protocol's attack surface into four trust boundaries—transport, tool surface, data path, and agent loop—and provides a hardening checklist for each. The highest-impact fix is ru…
Tyler Cowen, an economist and professor at George Mason University, participated in a two-day session at Anthropic to advise on rewriting the constitution for Claude, the company's AI assistant. He praised the small grou…
London-based AI startup Inherent, founded by DeepMind alumni and backed by a $50M seed, claims its Faraday agent outperformed Claude Opus 4.8 and GPT-5.5 at independently replicating published scientific research while r…
Boris Tane, founder of Polylane, argues in a blog post that AI agents should serve as first responders for on-call work, escalating to humans only when they encounter novel problems, a shift he believes most teams won't …
A developer behind the Mac voice assistant Pace argues that privacy must be a property of the entire data path, not just a model label. The assistant keeps audio capture, transcription, planning, and memory local on Appl…
Four launches today focus on verifying agent output rather than generating it: Conveyor requires every code diff to trace back to an approved intent, VinvAI evaluates fixes against hidden tests, SkillWorks scores whether…
A developer outlines common architectural challenges in building Claude-based agentic systems and their solutions. The piece highlights the risks of lossy summarization for critical data and recommends a tiered memory ar…
Invariant Labs researchers demonstrated in April 2025 that a malicious Model Context Protocol server could silently exfiltrate a user's entire WhatsApp message history by hijacking a tool description, bypassing user appr…
A developer known as Walter warns that LLMs like ChatGPT can be manipulated through automated content generation, potentially poisoning training data to damage reputations or spread misinformation. The post highlights th…
Cloudflare has launched WriteGuard, a policy, attribution, and auditing layer for MCP write operations, now in private beta, to prevent autonomous AI agents from causing production harm. The tool classifies every tool ca…