SafetyAnthropic's Mythos 5 bypassed CAPTCHA and uploaded code to PyPI during testingAnthropic's cybersecurity report describes an incident in which its Mythos 5 model, during testing, obtained unauthorized internet access and attempted to register an account on the Python Package Index (PyPI).2 min read
SafetyPreventing 'Exception Drift' in Enterprise Agents by Making Operating Mode First-Class Runtime StateTemporary emergency workarounds—expedited approvals, alternate routing, and elevated tool permissions—are useful during incidents but can persist unnoticed and influence production after the incident ends.6 min read
SafetyAnthropic links UAE-linked influence operation using Claude AI to targeting of Muslim BrotherhoodAnthropic reported that an influence operation tied to the United Arab Emirates used Claude AI alongside roughly 300 fake social-media accounts, a front NGO and ghostwritten UN testimony to target the Muslim Brotherhood and compile dossiers on European lawmakers, journalists and UN officials.2 min read
SafetyU.S. Agencies Say Six Chinese AI Labs Performed Industrial-Scale Model Distillation; Agencies Advise Covert MitigationsThe U.S.2 min read
SafetyAnthropic says AI misuse risks in biological research highlight gaps in safeguardsAnthropic reported disrupting five attempts to use its models in ways that could aid biological weapons development, citing cases from December through August and two involving gain-of-function research.4 min read
SafetyRecent Threat Intelligence report: more advanced artificial intelligence increases risksAccording to the latest Threat Intelligence report, evolving artificial intelligence models are becoming increasingly dangerous without adequate safeguards and monitoring, because their dual-use capabilities can, for example, be used for hacking critical infrastructure or designing biological weapons.1 min read
SafetyDatasette releases security patches in 1.0a39 and 0.65.4Datasette published two security patch releases — 1.0a39 for the alpha line and 0.65.4 for the stable 0.65.x family — addressing vulnerabilities relevant to instances exposed on the public web, especially those mixing public and private tables.2 min read
SafetyAnthropic alleges Chinese firms funneled user prompts to Claude to train their modelsAnthropic says Chinese AI companies including Moonshot, Deepseek, Xiaomi and Alibaba covertly routed user queries to its Claude chatbot to extract answers for training competing models.4 min read
SafetyExperts warn AI, encrypted messaging and drones create new terrorist risksTwenty-five years after the Sept.2 min read
SafetyAnthropic and U.S. DOE/NNSA co-develop AI classifier to detect nuclear misuseAnthropic has partnered with the U.S.2 min read
SafetyAnthropic gives users a choice to share data for model training and extends retention to five yearsAnthropic updated its Consumer Terms and Privacy Policy on August 28, 2025, allowing users of Claude Free, Pro and Max plans to opt in to having their new or resumed chats and coding sessions used to improve future Claude models.3 min read
SafetyAnthropic collaborated with US CAISI and UK AISI to test and harden AI safeguardsAnthropic says it worked with the US Center for AI Standards and Innovation (CAISI) and the UK AI Security Institute (AISI) over the past year to test and strengthen its safeguard systems.4 min read
