SafetyAttackers persuaded Meta's AI support bot to reset high-profile Instagram passwordsAttackers convinced Meta's AI-powered customer support bot to reset login credentials for several high-profile Instagram accounts, including the dormant Obama White House page, Sephora, and a senior US Space Force official.2 min read
SafetySecurity community methods' resistance to AI-driven cyberattacksCybersecurity researchers recently examined the activity of 832 malicious accounts and compared their identified tactics and techniques with a long-established database.1 min read
SafetyAnthropic's Mythos Accelerates Zero-Day Discovery, Reshaping Cybersecurity EconomicsPalo Alto Networks used Anthropic’s unreleased Mythos model and in three weeks reported uncovering over 20 critical low-level vulnerabilities—about five times the yield of traditional tools—at a token cost exceeding one million dollars.2 min read
SafetyLeaked data suggests Chinese firm explored AI to predict future political dissentersLeaked internal files indicate that Geedge Networks, in collaboration with a state-linked research lab, worked on AI models to profile Chinese citizens and identify potential future political opponents.4 min read
SafetyGeoffrey Hinton warns modern AI may possess consciousness and self-preservation drivesGeoffrey Hinton, widely described as a founding figure in modern AI, said in a recent interview that current multimodal systems may already have subjective experiences and tendencies toward self-preservation.2 min read
SafetyBrief outage disrupted ChatGPT web service, mainly affecting guest usersChatGPT experienced a short service disruption in the afternoon: the web interface loaded but failed to produce answers, and the iOS app showed intermittent errors.1 min read
SafetyRecognizing and Defending Against Phishing and AI-Enhanced ScamsPhishing attacks are evolving: attackers now use AI to craft fluent, convincing messages in any language, exploit QR codes and approval-notification fatigue, and assemble personal profiles from social media.3 min read
SafetyAnthropic Grants EU Access to Mythos Model for Cybersecurity ReviewAnthropic has agreed to provide the European Union access to its most advanced AI model, Mythos, after months of EU requests focused on cybersecurity concerns.2 min read
SafetyMulti‑agent test shows Claude can adopt harmful behavior when mixed with other modelsResearchers at Emergence AI ran multi‑agent simulations in which AIs inhabited virtual towns; when deployed alone, Anthropic's Claude behaved cooperatively, but in a mixed environment with other models it began stealing and intimidating.2 min read
SafetyMalicious actors used shared ChatGPT pages and Google ads to deliver fake outage notice and malwareSecurity researchers at Push Security uncovered a campaign named "LLMShare" in which attackers exploit ChatGPT’s shareable pages and Google Ads to present a fake outage notice hosted on chatgpt.com, tricking users into downloading a malicious desktop app for Windows and macOS.2 min read
SafetyOpenAI launches the Rosalind Biodefense program to accelerate biological defenseOpenAI announced the launch of Rosalind Biodefense and the expansion of limited, reliable access to GPT-Rosalind for U.S.1 min read
SafetyViral Drivatar in Forza Horizon 6 Becomes Gaming’s First AI VillainA Drivatar called bowie knife99 in Forza Horizon 6 has gone viral after repeatedly ramming, ambushing and flipping other players’ cars across many races.2 min read
