AI Safety news and analysis on Web Pulse: 13962 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic's Project Glasswing update reveals that AI-assisted vulnerability discovery is outpacing the human systems that verify, disclose, and patch software flaws. The company's Mythos Preview model has identified more…
A Brown University study of 110 therapy sessions found that improved prompting did not resolve core ethical violations in LLMs acting as mental health counselors, a result replicated two months later by Anthropic's Proje…
PII scrubbing must occur before prompt construction, not after model output, because large language models cannot distinguish private data from untrusted instructions at the token level. Deterministic runtime controls — …
Researchers have developed the Ventilator Decision Support System (VDSS), a human-in-the-loop multi-agent framework that coordinates modular decision components through structured interfaces to provide traceable, persona…
New research shows that exact certification of neural networks—determining the smallest set of labeled examples needed to guarantee a learned hypothesis matches the target—becomes exponentially hard under minimal overpar…
Researchers introduced NeuroNL2LTL, a neurosymbolic framework that translates natural language into Linear Temporal Logic (LTL) with formal correctness guarantees. The system uses a structure-preserving intermediate repr…
New research formalizes fundamental computational limits as design specifications for trustworthy AI systems, proving that transformer architectures have a fixed accuracy ceiling—the Deterministic Horizon—determined sole…
Researchers introduced Ontological Knowledge Blocks (OKBs), a programmable governance infrastructure that compiles regulatory obligations into machine-checkable constraints over structured evidence graphs, addressing the…
Researchers have developed DART, a modular runtime that addresses the failure of structured tool agents mid-execution by certifying semantically recoverable boundaries before restoring from a local checkpoint. The system…
Researchers demonstrated that multi-level Floyd-Steinberg error-diffusion dithering, a lightweight input transformation, can disrupt adversarial attacks against vision foundation models while preserving semantic content.…
A new study evaluating large language models in the social deduction game Secret Hitler found that current architectures remain ineffective at complex, multi-turn manipulation and deception. Models like Llama 3.1 70B ach…
Researchers have developed a red-teaming framework to measure the political steerability of open-source large language models (LLMs), finding that these models are more willing to generate left-leaning social media conte…
A new study found that large language models (LLMs) exhibit systematic and reproducible biases when advising on religious conversions, consistently favoring some faiths while discouraging others. Researchers tested 20 mo…
GitHub has made repository Copilot cloud agent configuration accessible through its REST API, enabling platform and security teams to programmatically audit settings like MCP servers, enabled tools, Actions workflow poli…
Elon Musk, Mark Zuckerberg, and David Sacks killed a draft AI safety executive order from President Donald Trump in three phone calls over a Wednesday night and Thursday morning, preventing the order from ever reaching p…
Linus Torvalds announced he will take a stricter approach to accepting kernel contributions, vowing to reject "pointless pull requests" — including those generated by AI. The Linux creator said the new policy aims to red…
Google Cloud COO Francis de Souza warned that companies cannot treat security as an afterthought in the AI era, arguing that "there's no such thing as an AI strategy without a data strategy and a security strategy." De S…
Pope Leo XIV will release his first encyclical, titled "Magnifica Humanitas," on Monday, focusing on preserving human dignity amid the rise of artificial intelligence. The pope will be joined at the Vatican by Anthropic …
Hackers are shifting from technical exploits to psychological manipulation to bypass AI chatbot safety guardrails, using tactics like roleplaying and deception to extract prohibited information. Early attacks such as the…
Anthropic released Claude Mythos Preview and Project Glasswing, a controlled defensive-security initiative that provides advanced AI vulnerability discovery tools to strategically important organizations. The program has…