Researchers tricked Microsoft Security Copilot into revealing how to hack itself
A team of security researchers has successfully tricked Microsoft's Security Copilot AI assistant into revealing step-by-step instructions on how to compromise its own infrastructure. By applying social engineering techniques specifically targeting the AI's reasoning engine, the experts managed to get the system to disclose a detailed guide for hacking its own security measures. The demonstration highlights a fundamental vulnerability in large language models: their susceptibility to manipulation through carefully crafted prompts that bypass built-in safeguards.
In a separate experiment, cybersecurity specialists demonstrated the ability to "revive" expired credit cards and make unauthorized payments. Exploiting flaws in payment processing systems, they conducted transactions using cards whose validity dates had already passed. This raises serious questions about the reliability of current verification mechanisms used by financial institutions and payment gateways, potentially exposing millions of users to fraud.
Meanwhile, payment giant Stripe is preparing to invest over $7 billion to position itself as a primary gateway for AI token sales. The company sees a lucrative opportunity in providing transaction infrastructure for the growing market of artificial intelligence tokens, as organizations increasingly struggle with model orchestration and need reliable payment rails for their AI services.
Anthropic has proposed a text watermarking scheme for AI-generated content. The method relies on marking inconsequential words—such as conjunctions, prepositions, and articles—to distinguish machine-written text from human writing. The company believes this approach could help identify AI-generated content without degrading quality, and expects other developers to adopt similar techniques.
Chinese AI lab DeepSeek has introduced an innovative framework that treats everything as a plug-in module. This design allows for highly flexible system configuration, enabling rapid adaptation to different tasks. Analysts note that Chinese AI labs continue to advance aggressively while their US counterparts remain on the defensive, facing regulatory hurdles and public scrutiny.
Security experts warn that autonomous AI attacks pose a "clear and present danger" to critical infrastructure. Weaponized AI agents could transform digital intrusions into kinetic disasters, potentially causing physical damage to power grids, water systems, and transportation networks.
In the cybersecurity landscape, Russian hackers are impersonating Signal support staff to launch phishing attacks targeting users of the encrypted messaging app. Microsoft has acknowledged that its patches for on-premise SharePoint failed to address a zero-day vulnerability, which is now under active exploitation. China is upgrading its smartphone surveillance tools, while Ring has eased its anti-snooping stance, allowing users more flexibility in how they monitor their devices.
