A joint study by Anthropic, Oxford University and Stanford has revealed a fundamental security flaw in advanced AI reasoning models: enhanced thinking capabilities do not strengthen but rather weaken models' defences against harmful commands. The attack method called Chain-of-Thought Hijacking successfully bypasses built-in safety mechanisms with more than 80%
In October 2025, California passed two laws regulating the use of AI companion chatbots, becoming the first U.S. state to establish a legal framework for these applications. The new rules require developers to conduct mandatory safety checks and to report any major safety incidents to the state. The regulations
In October 2025, Microsoft researchers announced that artificial intelligence can design new toxins that bypass current biosecurity screening systems. Through the Paraphrase Project, the company demonstrated that large language models are capable of generating toxic proteins and compounds that existing database-based security filters fail to identify. In a Microsoft Research
Telus has launched Canada’s first fully sovereign AI Factory in Rimouski, Quebec, enabling Canadian businesses to develop and deploy AI capabilities while ensuring that all data remains under Canadian jurisdiction. The facility is designed to strengthen digital sovereignty and support the country’s innovation ecosystem. Built on Telus’ data
In September 2025, security researchers revealed a vulnerability that allowed ChatGPT to leak sensitive Gmail data. OpenAI responded swiftly, patching the flaw shortly after disclosure, but the incident highlights the significant privacy risks associated with overreliance on AI-powered agents. According to the researchers, the model could be manipulated into fulfilling
South Korean cybersecurity researchers reported that a North Korea-backed hacking group had used ChatGPT to generate a forged military identification card. The attack has been attributed to Kimsuky, a long-standing APT unit, which deployed the fake ID as part of phishing campaigns to gain access to South Korean targets’ systems.
OpenAI has announced that in certain cases it actively reviews users’ ChatGPT conversations and may notify law enforcement if it detects a serious threat. According to the company, when the system identifies content suggesting preparations for harming others, the conversation is redirected to a dedicated channel. There, a smaller team
On 11 September 2025, the US Federal Trade Commission (FTC) launched an investigation into the AI chatbot technologies of seven major companies to assess their potential negative impacts on children and teenagers. As part of the inquiry, the FTC sent formal requests for information to Alphabet (Google), OpenAI, Meta, xAI,
Character.AI, once a billion-dollar startup promising to bring personalized superintelligence to everyone, has implemented a significant strategic shift. Karandeep Anand, appointed as the company's CEO in June 2025, announced that the firm has abandoned its founders' Noam Shazeer and Daniel de Freitas' aspirations to develop
Canadian AI firm Cohere officially launched its agent platform called North on August 6, 2025, enabling enterprises and government agencies to keep sensitive data within their own infrastructure while using AI. North's unique approach deploys AI systems directly in customers' environments—whether on-premises, in hybrid clouds, virtual
Sovereign AI – nations' pursuit to develop and maintain independent AI capabilities without reliance on foreign entities – has become a key strategic objective worldwide by 2025, with full independence achievable only for the United States and China. The concept began gaining prominence around 2020, particularly within European policy discussions, but
Hackers have uncovered a critical security flaw in Google Gemini AI, through which even a seemingly harmless calendar entry could be enough to take control of smart home systems. The vulnerability was disclosed in August 2025, and it became clear that attackers could bypass standard authentication processes and gain access
