AI Safety news and analysis on Web Pulse: 12756 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
The European Union Agency for Cybersecurity (ENISA) will meet with Anthropic on June 19, 2026, days after the US Department of Commerce ordered the company to suspend access to its advanced AI models for foreign national…
Sigil, an open-source tool for cryptographic prompt security in LLM applications, has been released. It provides tamper-evident audit trails and signed scopes without relying on external servers, using Ed25519 signatures…
A security researcher discovered a prompt injection vulnerability in Firefox's AI chatbot integration in October 2025, allowing attackers to steal personal information such as emails via malicious page titles. The flaw e…
Prime Minister Narendra Modi urged G7 leaders in Evian, France, to adopt a human-centric approach to AI development, emphasizing the need to empower ordinary people and protect children from misinformation and deepfakes.…
A developer with 20 years of experience warns that heavy reliance on AI tools leads to skill atrophy, where fundamental problem-solving abilities weaken over time. The article highlights how AI-generated solutions can pr…
Kansas City, Missouri, plans to install facial recognition cameras on public buses to identify banned riders and missing persons, sparking debate over security versus privacy. The American Civil Liberties Union warns the…
EPFL researchers released MeditronFO, the first fully open framework for building medical large language models, making every stage of development publicly available to ensure transparency and accountability in AI health…
Researchers introduced CaVe-VLM-CoT, a modular reflection-based agentic-RAG framework that enforces evidence-grounded reasoning in vision-language models through a five-stage closed-loop pipeline. The framework achieves …
Researchers introduced DeFAb, a benchmark for defeasible abduction in foundation models, converting knowledge bases into 372,648+ logically verifiable instances. Frontier language models achieved at most 65% accuracy, dr…
Researchers introduced ForecastBench-Sim, a simulated-world forecasting benchmark built on game rollouts from Freeciv, to evaluate AI systems' probabilistic reasoning under dynamic world states. The benchmark generates c…
Researchers introduced Xcientist, a research harness that externalizes research synthesis and experimental validation into inspectable, contract-governed processes. The system addresses claim drift in automated research …
Researchers introduced SciRisk-Bench, a benchmark evaluating AI safety in scientific contexts across 7 disciplines and 10 risk dimensions. Tests on mainstream and science-oriented LLMs revealed where models remain unsafe…
Researchers propose a fully local AI cascade for de-identifying educational dialogue that achieves 0.958 macro F1 on math tutoring transcripts, outperforming a commercial API (0.706) and LLM-only baselines (0.767), while…
Researchers at an undisclosed institution have found that large language models (LLMs) poorly preserve diagnostic uncertainty in clinical text, altering phrases like "possible pneumonia" more than half the time, which co…
Researchers found that unlearning knowledge from large language models causes collateral damage that decays with semantic distance but persists across domains. They developed a method to audit this damage before unlearni…
Researchers introduced TRIDENT, the first multi-agent reinforcement learning framework that co-designs three components to cancel biases from hybrid discrete-continuous actions, safety constraints, and physics-governed d…
Researchers propose SAGE, a post-hoc method to sanitize the final update vector in LLM unlearning, reducing the trade-off between forgetting and retention without rerunning the original pipeline. SAGE uses spectral activ…
Researchers introduced Self-Consistency Training with Reinforcement Learning (Self-CTRL), a method that aligns language models' self-explanations with their actual behavior. In tests, the approach improved correlation be…
Researchers introduced Forged Calamity, a benchmark dataset of 30,000 images for detecting synthetic disaster scenes generated by diffusion models. Experiments showed current forensic detectors lose up to 50% accuracy on…
Snyk is offering its full AI Security Platform free to open source maintainers through the Secure Developer Program, providing risk-based prioritization and automated remediation to help them manage vulnerability backlog…