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.
Anthropic reported that its Project Glasswing initiative, launched in April to use its Mythos AI model for identifying software vulnerabilities, has detected 23,000 potential flaws across 1,000 open-source projects. Desp…
Anthropic's Claude Mythos program has reported 23,019 vulnerabilities but only demonstrated fixes for 75, raising questions about its claimed frontier-exclusive capabilities. Independent teams have reproduced the showcas…
The White House's executive order on AI governance signals a shift toward industry-government collaboration, but effective oversight requires accountability rather than heavy-handed regulation. Anthropic's responsible re…
A new concept called "Judgment Compression" is emerging as a critical missing layer in production AI system design, addressing the gap between generating outputs and validating their structure. While validators can confi…
AI scribes are being rapidly adopted in clinical settings, but their accuracy can drop to 65% in complex cases, and they may introduce bias or privacy risks through hidden transcripts. Clinicians are urged to review note…
Fly.io developers are building AI agents that avoid self-destruction by executing risky commands in ephemeral sandboxes called Sprites, keeping the agent loop separate from code execution. This architecture, demonstrated…
The EU AI Act imposes board-level accountability for AI governance, requiring CIOs, CISOs, and compliance leaders to establish organizational structures such as AI inventories, designated human oversight, and documentati…
Microsoft's GitHub Copilot shifts to per-token billing, signaling the end of subsidized AI and passing real costs to consumers and enterprises. Open-source AI SRE tool Nightwatch (ninoxAI) launches with read-only agentic…
A developer warns that AI-generated tests often contain 'rotten green tests' that pass without actually verifying anything, citing research showing up to 62% of LLM-generated tests fail to compile and that AI-co-authored…
Om Malik critiques Anthropic for naming its most powerful AI model Mythos, arguing the term encodes an instruction to receive rather than verify, contrasting with the company's stated mission of AI safety. He draws paral…
AI security defenses built for chatbots fail against prompt injection in agentic systems because agents have multiple input channels—documents, tool outputs, memory—that are not monitored or filtered. Unlike chatbots wit…
Hackers on the Hill returns to Washington DC on June 16, 2026, as the first Capitol-side researcher-to-policymaker gathering of the year. The White House signed the Mythos-era AI cybersecurity executive order on June 2, …
Congress has failed to pass comprehensive tech regulation for decades, while states like California are moving to regulate AI risks such as privacy violations, job losses, and existential threats. AI leaders Sam Altman a…
A reader reported reading fewer articles than usual in week 23 of 2026, citing time spent catching up on recorded talks from re:publica 26 and personal struggles with depression. Notable reads included an analysis of Cha…
A developer describes running autonomous AI agents that merge dozens of pull requests daily across six repositories without human intervention, but argues that commit logs and screenshots cannot prove the process is real…
An injured teenage survivor of a January 2025 school shooting in Nashville, Tennessee has sued Omnilert, the manufacturer of an AI gun detection system that failed to detect the handgun used in the attack. The lawsuit, f…
A writer argues that artificial intelligence models are increasingly sophisticated statistical tools that mimic human output without genuine understanding or embodiment, warning that optimizing for testable properties le…
The Atlantic published an essay titled "Welcome to the Slopverse" on November 26, 2025, arguing that generative AI produces multiple coexisting realities rather than simply "hallucinating." The piece, part of a series ma…
A new benchmark dataset, the Social Media Robustness Benchmark, evaluates how deepfake detectors perform on images re-encoded by social media platforms like Instagram, Facebook, TikTok, and X. The dataset contains 2,400 …
In a 2025 benchmark comparing three industry static analysis tools against three frontier LLMs on sixty-three real vulnerabilities in ten C# projects, the LLMs achieved F1 scores between 0.75 and 0.80, significantly outp…