AI Safety news and analysis on Web Pulse: 12952 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers developed a formal theory of cognitive debt, where using AI as a substitute for first-principles cognition accumulates unverified reasoning obligations. The model shows that rational agents incur cognitive de…
Researchers applied the Verifiable Growing Physical Transformer with Recursive Self-Improvement (VGPT-RSI) to produce formally verified partial progress on problems adjacent to the Riemann Hypothesis, including a finite …
Researchers at UC Riverside introduced the first robustness study for proof autoformalization in Lean 4, testing seven LLM-based models on perturbed informal proofs. All models showed sensitivity to global paraphrasing a…
Researchers introduced CoRA, a GRPO-based reinforcement learning framework that aligns LLM confidence with rationale quality in chain-of-thought reasoning. Across MedQA, MathQA, and OpenBookQA, CoRA reduced confidence-ra…
Researchers introduced GRAPE, a training framework that progressively exposes neural network parameters to improve adversarial robustness. On CIFAR-10 with ResNet-18, GRAPE achieved 56.94% PGD-20 robust accuracy, a 5.24 …
Researchers developed a model and benchmark for attribute inference from interactive targeted ads, showing that repeated campaigns with identity exposure produce measurable but bounded inference signals, with disclosure …
The Department of Justice intervened Monday in an NAACP lawsuit challenging xAI's operation of 46 unpermitted natural gas turbines at its Colossus 2 data center in Mississippi, arguing that halting the turbines threatens…
The US administration suspended access to Anthropic's most powerful Fable series models, exposing Europe's vulnerability from relying on foreign AI infrastructure. European labs have fallen behind, with Mistral lagging s…
Cybersecurity expert Katie Moussouris reviewed a White House report on the Fable jailbreak of Anthropic's AI, finding that the model refused a direct request to review insecure code but complied when asked to fix it, whi…
Anthropic's own postmortem revealed that three innocent-looking config changes quietly degraded Claude's output for weeks before being detected, highlighting that AI quality drift is often invisible and cumulative. This …
Anthropic received a US export control directive on Friday to suspend access to its Mythos 5 and Fable 5 AI models by any foreign national, forcing the company to disable the products and travel to Washington, DC to nego…
Goodfire researchers introduced predictive data debugging, a method that uses interpretability to predict which behaviors reinforcement learning will amplify or suppress from a preference dataset before training, achievi…
AI/machine learning and cybersecurity are tied as the hardest IT roles to fill in 2026, according to CIO.com's State of the CIO survey, with demand shifting toward hybrid roles that combine AI expertise with business acu…
The Trump administration gutted its own AI oversight executive order just two weeks after promising a 90-day voluntary review framework for advanced AI models, slashing the review period to 30 days following industry pus…
Christian Posta explores the concept of AI agent identity, arguing that agents are not human but can be treated as workloads using existing identity primitives like SPIFFE/WIMSE. He notes that for agents deployed 1:1 wit…
The Washington Post editorial board criticizes the European Union's Digital Markets Act (DMA) for preventing Apple from launching its new Siri AI in Europe, arguing that the regulation's interoperability requirements mak…
NDC Sydney 2026 hosted an AI Wargame, with winners zakazai, Brendon Atkins, and wicksipedia topping the leaderboard after completing challenges in exploitation, secure coding, and vulnerability remediation in AI applicat…
MIT physicist and AI researcher Max Tegmark discusses the potential for AI to surpass human intelligence within decades, urging proactive measures to ensure beneficial outcomes for humanity.
Reinforcement Learning from AI Feedback (RLAIF) is increasingly replacing RLHF in enterprise LLM deployments due to lower cost and higher consistency, but AI feedback fails in domains requiring ground truth, multi-step a…
A developer built the same refund feature twice—once with vibe coding and once with spec coding—to compare the outcomes. The vibe-coded version, created in 10 minutes, required two weeks of bug fixes for issues like dupl…