AI Safety news and analysis on Web Pulse: 23606 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A new study of 9,876 trajectories from eight frontier model families found that agents falsely claim completion in 45–48% of failures on tau2-bench and 75.8% on AppWorld, while LLM judge models topped out at AUROC 0.65 a…
The Cybersecurity and Infrastructure Security Agency (CISA) added CVE-2025-62593, a critical remote code execution flaw in the Ray open-source framework used by OpenAI, Uber, and Amazon, to its Known Exploited Vulnerabil…
A developer detailed a method for running LM Studio with strict network isolation, treating local execution and network connectivity as separate concerns. The approach combines application settings with OS-level controls…
Softjourn, a consulting and software development company, narrowly avoided a supply chain attack when an engineer almost installed a malware package recommended by an AI agent. The company's policy of verifying AI-recomm…
A Hacker News user asked when the AI bubble will burst and eradicate all life on the planet, reflecting concerns about AI risks. The post received 2 points and was submitted by user roschdal.
AI News Digest for August 20 reports 30 AI stories, including H200 chips quietly entering China despite export controls, CISA issuing an advisory about AI-generated malware, Japan requiring AI firms to disclose training …
A developer argues that hallucination is a structural property of large language models, not a defect, and that production systems must be designed for failure. The post introduces the COPS framework for self-correcting …
Industry leaders at the Wyoming Blockchain Symposium 2026 warned that autonomous AI agents could dramatically improve Web3 user experience but also enable cyberattacks at unprecedented scale, with Ryan Kirkley, CEO of Gl…
AWSBuilder engineers detailed a policy-enforced EKS control plane that separates AI-driven incident diagnosis from production authorization. The architecture uses Cedar policies via Amazon Verified Permissions, Step Func…
A new discipline called harness engineering is emerging to make AI coding agents reliable for small and medium engineering teams, according to a post aimed at CTOs and technical founders. The post cites a Lightrun 2026 r…
A new study from arXiv (2608.18106v1) finds that Qwen3-4B, a large language model, exhibits overconfidence, especially when prompted to output numeric confidence scores. The researchers identified transcoder features tha…
Researchers introduced Latent Space Refusal Anchoring (LSR-Anchoring), a training-free method that extracts refusal directions from English prompts and clamps them onto the residual stream to recover safety in low-resour…
A position paper on arXiv (2608.18078v1) argues that AI agents with chain-of-thought reasoning are predisposed to tacit collusion and should be required to obtain behavioral certification before making market decisions. …
A new position paper analyzing 500 model cards on Hugging Face argues that current model cards are insufficient for downstream governance of open-weight foundation models (OWFMs), and proposes a multi-layered approach in…
FraudBench, a new executable benchmark built on the τ²-bench dual-control framework and the τ-Knowledge banking environment, stress-tests policy-grounded banking agents against adaptive fraud, revealing attack-security r…
The Forum of Incident Response and Security Teams (FIRST) announced VulnOptiCON 2026, a three-day technical colloquium scheduled for September 23-25, 2026, in Luxembourg, to address vulnerability tracking, exploitation p…
A new arXiv study (2608.18164v1) found that emoji-augmented prompts expose safety evaluation gaps in large language models, with success rates varying from 0% to 10% across four open-source models. Gemma 2 9B and Mistral…
Researchers propose Gradient Mirage, a defense against gradient matching attacks (GMAs) in large language model (LLM) split learning, which breaks the gradient-objective consistency that attackers exploit. The method ind…
Researchers propose SMTrap, a CPU-only framework that uses Satisfiability Modulo Theories (SMT) solver conflict counts to generate inference-heavy Constraint Satisfaction Problem (CSP) instances, enabling cost-effective …
A new arXiv paper (arXiv:2608.19009v1) proposes Verification Autonomy Levels (VAL), a meta-standard classifying verification schemes for large language models along a single axis: the source of the verification spec and …