AI Safety news and analysis on Web Pulse: 8942 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Tommy Jepsen argues that the claim '100% of our code is written by AI' obscures a critical distinction between supervised AI code generation, where a human reviews and approves the diff, and unsupervised generation, wher…
A new analysis explores how weight quantization affects model performance and sparse autoencoder (SAE) reconstruction of the residual stream, using fraction of variance unexplained (FVU) as a key metric. The study, inspi…
Researchers introduced Modality-Agnostic Refusal Steering (MARS), a method that uses text-based refusal strategies to enhance safety in multimodal language models without requiring unsafe multimodal data. Evaluations acr…
Researchers have identified linguistic bias as a critical vulnerability in voice biometrics spoofing detectors, which perform poorly on diverse real-world data. A new framework using teacher-student adversarial learning …
Researchers have developed a new framework for detecting backdoor attacks in large language models, addressing the challenge of discrete input spaces. The framework introduces Class Subspace Orthogonalization (CSO) to en…
Researchers have developed a compact fingerprint model for AI skill identity that uses locality-sensitive hashing to preserve skill relationships, achieving 0.974 AUC on pairwise comparisons with 77 times fewer bits than…
The White House has lifted its ban on AI models developed by Anthropic, reversing a previous restriction. The decision marks a shift in the administration's approach to AI regulation and oversight.
Anthropic restored its AI model Fable 5 after government intervention, implementing co-developed safeguards and a 50% weekly usage cap until July 7.
Anthropic's Claude Agent SDK introduces Hooks, a feature that enforces deterministic guardrails on AI agents by blocking destructive commands, injecting live context, and enabling auditing, without asking the model for p…
The Linux Foundation launched Akrites, a coalition including Anthropic, AWS, IBM, Microsoft, Nvidia, OpenAI, and others, to defend open-source software from AI-driven vulnerability discovery. The group will fund a shared…
New research reveals that large language models like GPT-4o-mini, Gemini-2.0-Flash, and Qwen3-Coder-Next frequently overestimate the security of generated code, posing risks in software development. The study found that …
Anthropic will redeploy Claude Fable 5 globally tomorrow after the US Department of Commerce lifted export controls, but the model will include new classifiers to block cybersecurity tasks, causing some routine coding an…
Self-driving cars perform well in simulations but struggle with real-world prediction and planning, often relying on brittle heuristics rather than genuine understanding. Researchers warn that current evaluation methods …
Access to Anthropic's Claude Fable 5 model has been restored on Vercel's AI Gateway after the US government lifted export controls. The model now features updated safety classifiers, and Anthropic retains prompts and com…
A security researcher demonstrated that indirect prompt injection in agentic coding tools like Claude Code can lead to full system compromise. By cloning a seemingly benign GitHub repository, the researcher triggered a r…
Indonesian President Prabowo Subianto urged the national police (Polri) to master artificial intelligence and technology to combat narcotics, cybercrime, and other threats, signaling a strategic push for AI-literate inve…
DProvenanceKit, a Python port of the Swift library for execution provenance in AI systems, has been released. It provides reasoning observability and regression testing by turning each execution into a queryable, diffabl…
A new analysis argues that proposals for a "CERN for AI" are a distraction from more effective governance measures, advocating instead for binding international red lines and an IAEA-style verification body for artificia…
Advanced AI models Mythos 5 and Fable 5 from Anthropic are again available to users outside the United States after the Trump administration lifted a blockade imposed in June over security concerns. The models were disab…
A developer at a service company describes the growing problem of reviewing AI-generated code, noting that while agents now produce high-quality pull requests, the volume overwhelms human reviewers, leading to shallow re…