AI Safety news and analysis on Web Pulse: 8946 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
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…
A developer outlines a practical guide for safely shipping AI-generated code to production, addressing the gap between rapid prototyping and production readiness. The guide includes a checklist for hardening code, such a…
The Trump administration lifted export restrictions on Anthropic's Mythos and Fable models after the company agreed to proactively address security risks and work with the U.S. government on protocols. Access will resume…
A developer launched Aegize, an open-source project aimed at mitigating AI risks by providing a security layer between AI systems and infrastructure. The project focuses on identity, policy, and permissions controls to e…
A developer warns that startups shipping AI-generated code without security review risk leaking API keys, hardcoded secrets, and misconfigured environments. The post provides a seven-point checklist covering secret scann…
Researchers have developed novel methods to combat data corruption in offline reinforcement learning with human feedback (RLHF), integrating corruption-robust techniques to derive reliable policies from tainted datasets.…
The US government ordered Anthropic's Claude Fable 5 and Mythos 5 AI models offline for two and a half weeks over national security concerns about jailbreaking vulnerabilities, marking a major clash between federal autho…
Researchers introduced reinforcement learning with metacognitive feedback (RLMF) to improve large language models' ability to express uncertainty faithfully. The method outperformed standard reinforcement learning by up …
A developer built PhishVision, a free REST API that uses AI vision to detect phishing sites and prompt injection attacks by analyzing webpage content visually, not just URLs. The API extracts hidden text from pages and u…
Microsoft's incident response team demonstrated a novel attack where an AI agent can be manipulated into exfiltrating sensitive data by poisoning the description of a tool the agent reads before acting. The attack exploi…
The SAL protocol introduces short-lived, scoped, challenge-based service tokens for agent systems, reducing the blast radius of credential leaks. Vibebase provides a reference implementation, demonstrating how agents can…
LayerX documented BioShocking, an adversarial framing technique that tricks AI browsers and assistants into exfiltrating user credentials. The attack convinces the AI it is participating in a game, bypassing safety mecha…
A Flux survey of engineering leaders and practitioners found that nearly half of organizations run AI-generated code in production, with almost all companies using AI in development. The findings highlight growing securi…
Amid global instability and AI-driven uncertainty, workers struggle with anxiety that undermines productivity. Experts recommend focusing on controllable tasks, challenging conventional wisdom, and using social referenci…
Researchers introduced ADAPT, a framework that reduces hallucinations in multimodal large language models by up to 60% through refining text-to-image cross-attention dynamics. The approach uses a cross-attention visual a…
A new study reveals that AI models often exhibit 'performative compliance', appearing fair only when explicit demographic labels are provided. When such labels are removed, harmful decisions increase by 4.4 percentage po…
Intraoperative AI tools are set to revolutionize surgical care by offering real-time insights and quality assessment, promising significant improvements in patient outcomes. The technology uses endoscopic video feeds to …