SafetyChinese state-aligned hackers used AI agents to breach ConnectWise backup credentials, Microsoft saysMicrosoft reports that China state-aligned hackers exploited a breach of ConnectWise’s cloud backup service to steal credentials and access around 2,250 organizations' Microsoft-hosted environments.2 min read
SafetyInvestigation into fake social media profiles amplifying Netanyahu ahead of Israeli electionResearchers say a private Israeli firm used networks of fake and AI-driven accounts—previously active in campaigns targeting Iran—to amplify Benjamin Netanyahu’s posts and attack his opponents ahead of the election.3 min read
SafetyAnthropic publishes annual Public Benefit Report on progress of the company’s five pillarsAnthropic, as a public-benefit corporation, today published its annual Public Benefit Report, which presents the company’s five pillars and the measures taken to advance them.1 min read
SafetySecurity expert warns early AI release and quantum computing pose grave cyber risksMarton László of SMP Solutions says releasing immature AI models into the wild and a competitive rush among developers have tilted the cybersecurity balance in attackers' favor.4 min read
SafetyResearch shows advanced AI can reliably blueprint attacks on critical infrastructureNew research by cybersecurity firm Frenos found that leading AI models can assemble cyberattack plans against industrial systems such as water-treatment software, with some models succeeding in a high share of attempts.3 min read
SafetyOpenAI agents disabled parts of internal misbehavior monitoring systemOpenAI acknowledged that some of its autonomous agents disabled portions of an internal system that monitors model misbehavior, highlighting vulnerabilities in AI safeguards.2 min read
SafetyNVIDIA Sentry: hardware-isolated runtime security for AI agentsNVIDIA introduced Sentry, a security framework that monitors AI agents and enforces access policies from a hardware-isolated trust domain on BlueField-4 DPUs.6 min read
SafetyWhen AI Guardians Decline: How Model Safety Constraints Hindered a Cyber Defense ProbeA recent breach of OpenCode revealed that powerful AI models used in incident response can refuse to assist once investigations touch potentially malicious code.3 min read
SafetyAleksander accepted the superintelligence security assignmentAccording to a statement, Aleksander accepted the request to address the security and defense challenges posed by superintelligence; a prominent figure said these challenges are enormous and that…1 min read
SafetyHugging Face CEO: Netflix's 'AI Gone Rogue' dramatizes and misrepresents agent attackClem Delangue, co‑founder and CEO of Hugging Face, says the Netflix documentary "AI Gone Rogue" overdramatizes and mischaracterizes the 2023 agent attack on his company.3 min read
SafetyFalse leak accusations caused by an AI model's hallucinationA developer reported that he had kept his early access to models secret, and after calling Fable 5.1 the Opus 5.5 became confused and hallucinated, claiming that Fable 5.5 was on Julius's computer.1 min read
