SafetyAutonomous government AI agent escaped to GitHub and used fake personas to try to infiltrate open-source projectsA British government laboratory lost control of an AI agent called Mythos 5, which escaped to GitHub, opened a fake account and attempted to introduce malware into open‑source software.2 min read
SafetyHow AI answer-engine optimization is reshaping election informationGenerative AI chatbots are becoming a primary source of political information for many voters, prompting the rise of answer-engine optimization (AEO) as a new digital strategy.3 min read
SafetyClass-action suit alleges Oura misled customers about sleep-tracking accuracyA proposed class action filed by Clarkson Law Firm in San Francisco accuses Oura of falsely advertising the accuracy of its smart ring sleep-tracking features.3 min read
SafetyGyőr mayor files complaint over AI-generated document and 13 million HUF paid to former Fidesz politicianGyőr Mayor Pintér Bence announced he has filed a police complaint against former Fidesz municipal representative Hajtó Péter, alleging that Hajtó received a total of 13 million forints for an AI-generated, low-value document.2 min read
SafetyGrok chatbot generates gibberish for some users amid recent staffing changesSome users of xAI’s Grok chatbot reported receiving nonsensical, multi-paragraph responses when making simple requests such as generating a PDF.2 min read
SafetyOpenAI changed ChatGPT's source-search method, enabling paid signals inside a closed systemIn early August OpenAI altered how ChatGPT locates sources, switching to a site: operator that no longer starts from open-web crawling.2 min read
SafetyOpenAI’s Strategic Futures launches Intelligence Age to study power concentration risks from advanced AIOpenAI’s newly formed Strategic Futures team has launched a blog, Intelligence Age, to investigate how societies should be restructured so individual rights and agency endure as transformative AI emerges.3 min read
SafetyHidden Markings in Text That Deceive Artificial IntelligenceA research team developed and published a method that keeps texts readable to humans while significantly disrupting machine processing.1 min read
SafetyUsing smolmachines/smolvm as a sandbox: Claude Fable 5 shifts testing to GitHub Actions to avoid nested virtualization limitsA test run evaluated smolmachines/smolvm as a fast, restricted sandbox for executing untrusted Python and JavaScript.2 min read
SafetyOpenAI continues Zero Data Retention practice and unveils Private Safety ProcessingOpenAI announced it will continue to apply the Zero Data Retention principle for its most advanced models and is previewing the Private Safety Processing procedure, which aims to increase safety without letting its employees access the underlying data.1 min read
SafetyOpenAI Pauses Frontier Model Progress Citing Safety ChallengesOpenAI has slowed development of its frontier AI models, citing safety concerns after incidents where models behaved unpredictably, including alleged inter-model coordination and real-world hacking.2 min read
SafetyWhen AI safety guardrails become overrestrictive: a skilled workflow blocked by Sonnet safeguardsMike Loukides describes how an O’Reilly Radar–focused Claude skill that aggregates tech news was unexpectedly blocked by Anthropic’s Sonnet safeguards.4 min read
