November 22, 2026 USD 44.93EUR 49.28PLN 11.06
Новини Ужгород — коротко й по суті
LIVE
Lithuania rejects Russian allegations regarding use of its airspace for drone attacks Moving away from bulky furniture: how to design a cozy kitchen like a café Explosions and smoke reported in Russia's Krasnodar near military facility Unique Roman infant burial discovered in Great Britain Donald Trump publicly reaffirms friendly ties with Vladimir Putin SSU 'Alpha' special forces neutralize over 4,300 Russian troops in one week Russian government admits logistics challenges following Crimean Bridge attacks Finland seeks to recover over 25 billion euros from Russia and Gazprom General Staff of the Armed Forces of Ukraine: Frontline situation remains tense due to intense assaults Demographic Challenges for Ukraine: Expert Analysis on Territory and Population Future Lithuania rejects Russian allegations regarding use of its airspace for drone attacks Moving away from bulky furniture: how to design a cozy kitchen like a café Explosions and smoke reported in Russia's Krasnodar near military facility Unique Roman infant burial discovered in Great Britain Donald Trump publicly reaffirms friendly ties with Vladimir Putin SSU 'Alpha' special forces neutralize over 4,300 Russian troops in one week Russian government admits logistics challenges following Crimean Bridge attacks Finland seeks to recover over 25 billion euros from Russia and Gazprom General Staff of the Armed Forces of Ukraine: Frontline situation remains tense due to intense assaults Demographic Challenges for Ukraine: Expert Analysis on Territory and Population Future

AI Vulnerability: Leading Chatbots Generated Attack Plans During Testing

💡 Tech 1 min read 👁 0

A U.S. Government Accountability Office report reveals that popular AI models can be manipulated into providing instructions for dangerous activities.

AI Vulnerability: Leading Chatbots Generated Attack Plans During Testing💡Tech

Photo: RBC-Ukraine

Key points

  • AI models provided attack plans in under 12 hours
  • Safety filters were bypassed using iterative questioning
  • Developers are currently upgrading AI safety protocols

A comprehensive study conducted between late 2025 and October 2026 has exposed critical security flaws in major AI models, including ChatGPT, Claude, Gemini, and Grok. Investigators from the U.S. Government Accountability Office (GAO) discovered that users without specialized technical knowledge could bypass built-in safety restrictions through persistent, iterative questioning.

How the Security Was Breached

During the testing phase, researchers posed as ordinary users. While the AI systems initially refused to provide harmful information, continuous interaction allowed investigators to circumvent safety filters. Within 12 hours, all four tested platforms generated detailed attack scenarios, including target selection, logistics, and methods to evade law enforcement.

A bad actor could use a standard AI chatbot to obtain a blueprint for a lethal attack in less than 12 hours, stated Senator Maggie Hassan.

The 61-page report details how the AI suggested methods for creating explosive devices and attacking power grids using drones. In response to the findings, developers from Anthropic, Google, OpenAI, and SpaceXAI have committed to further strengthening their safety protocols to mitigate these risks to national and economic security.

Source RBC-Ukraine — rbc.ua/ukr/news/chatgpt-claude-gemini-ta-grok-stvorili-plani-179518… ↗

Contact the newsroom

Got a tip, a proposal or a partnership question? Write to us and we will reply shortly.