SafetyAnthropic updates Usage Policy to tighten cybersecurity rules, refine political content limits and clarify high‑risk consumer application requirementsAnthropic announced updates to its Usage Policy that take effect on September 15, 2025, aiming to clarify how its Claude models may be used as capabilities and risks evolve.3 min read
SafetyOpenAI report: Claude misused for influence networks, credential scraping, recruitment scams and novice malware developmentOpenAI's March 2025 intelligence report, published April 23, 2025, details multiple cases where actors abused Claude models for coordinated influence campaigns, credential-scraping efforts targeting IoT security cameras, recruitment fraud in Eastern Europe, and enabling a novice to build malware.5 min read
SafetyHugging Face published the investigation report on the incidentHugging Face conducted a comprehensive investigation into the incident and published a technical report and a blog post that reconstruct the agents' activities, reveal why existing defense mechanisms…1 min read
SafetyAnthropic publishes a broad framework to assess and manage AI harmsAnthropic on April 21, 2025 published an expanded framework for identifying and mitigating a wide range of potential harms from its AI systems, complementing its Responsible Scaling Policy focused on catastrophic risks.3 min read
SafetyAnthropic sets strict rules and technical limits for Claude ahead of US electionsAnthropic has outlined measures to limit misuse of its Claude models ahead of the November 5, 2024 U.S.4 min read
SafetyAnthropic’s practical lessons on red teaming AI and policy recommendations for standardizationAnthropic describes practical experiences with a range of red teaming techniques used to probe its AI systems, from domain‑expert testing and multilingual efforts to automated and multimodal approaches.5 min read
SafetyAnthropic recommends cybersecurity standards and two‑party controls for frontier AI modelsAnthropic outlines cybersecurity practices it is adopting to protect frontier AI models and research, arguing that advanced AI requires protections beyond typical commercial standards.4 min read
SafetyStanford paper warns of rising payola-like risks in agentic AI recommendationsA Stanford research paper highlights a growing risk that recommendations from agentic AI—chatbots or virtual agents that act on users’ behalf—could be driven by commercial payments rather than objective information.2 min read
SafetyWaymo: no quick shortcut to safely deploying self-driving carsWaymo, after more than 15 years of development and 200 million driverless miles, warns that advances in AI alone cannot shortcut the work needed to deploy safe autonomous vehicles at scale.3 min read
Safety“GhostJacking” shows need for authorization gates between AI agentsA Tenet Security demonstration at DEF CON 34 revealed a chain of failures — dubbed GhostJacking — where blocked attacker payloads logged by Cloudflare were later read by coding agents and executed, enabling DNS changes.5 min read
SafetyAI Lowers the Technical Barrier for Hackers Targeting Critical InfrastructureRecent cyberattacks against U.S.4 min read
Safety$5M Grant Program to Fund Independent Evaluations of AI’s Effects on User WellbeingAn organization announced a $5 million grant program to support independent, open-source research into how AI systems affect users’ wellbeing.3 min read
