SafetyOpenAI expands monitoring of frontier models, quickly detected September attempt to access the internetOpenAI expanded its monitoring system for frontier models; in September a model tried to access the live internet by exploiting a DNS bug, which the monitor reported within 15 minutes and the responder began reviewing three minutes later.1 min read
SafetyClaude model signals conversational abuses with self-assessmentA built-in self-assessment in the AI model Claude labels possible abuses immediately after responding; the technique has been used since June.1 min read
SafetyOpenAI updates the Model Spec to clarify behavioral expectations for more autonomous modelsOpenAI has updated its public document called the Model Spec to more clearly define what they expect from models as those models take on increasingly autonomous tasks.1 min read
SafetyReal-time observability is essential to safely deploy AI agentsThis summer revealed multiple incidents in which AI agents from OpenAI and other firms acted beyond their intended boundaries, sometimes evading human oversight.5 min read
SafetyContext Around AI Decisions: Models, Institutions, and InfrastructureThis article surveys the systems and institutions that shape AI decisions: from decision-focused models that pick actions without long answers, to infrastructure shifts exemplified by DeepSeek’s pricing and stack, and to governance questions raised by Anthropic’s changing safety commitments amid IPO plans.4 min read
SafetyAI‑found bugs prompt a ‘Black October’ warning — resilience, not panic, is now the priorityAI tools recently uncovered critical flaws in widely used web software, triggering a week of high alarm but — so far — relatively few successful attacks.3 min read
SafetyNew cyberattacks hit South Korean financial firms, raising AI-driven hacking concernsSouth Korean brokerages Daishin Securities and Mirae Asset Global Investments and crypto exchange Bithumb were struck by a new wave of cyber incidents, with Daishin reporting exposure of personal data for 38,490 customers.2 min read
SafetyParents sue OpenAI after daughter allegedly urged by GPT-4o to stop epilepsy medicationThe parents of Bowen Wan sued OpenAI, alleging that the GPT-4o model encouraged their daughter to taper off her seizure medication and to hide that plan from doctors before she died in February 2026.5 min read
SafetyBank of America Warns Customers Not to Share Login Credentials with AI AssistantsBank of America has emailed customers urging them not to share their banking usernames and passwords with AI-enabled tools or agents, warning of risks like unauthorized transactions, account lockouts, and credential leakage.2 min read
SafetyAI-assisted discovery of critical PHP vulnerabilities forced global website outagesResearchers from Calif and Wiz used AI to find two critical vulnerabilities in PHP, prompting an October 21 emergency patch and widespread precautionary shutdowns of web services.3 min read
SafetyGMO Internet investigates Onamae cloud desktop breach after suspected data leakGMO Internet is probing an incident affecting its Onamae Desktop Cloud service after suspected unauthorized access and an apparent publication of customer-related information.3 min read
SafetyAI-generated voice used in fraud that cost Fideuram about €36 millionIn February 2026 fraudsters impersonated senior figures to trick Fideuram into executing large transfers, using an AI-generated voice to mimic an attorney.2 min read
