November 22, 2026 USD 44.93EUR 49.28PLN 11.06
Новини Кременець — коротко й по суті
LIVE
New Energy Strikes: Russia Attacks Naftogaz Facilities and Filling Stations Slovakia records highest bustard population in over 50 years China completes bridge construction over 400-meter canyon Swedish startup develops new K200 interceptor drone against jet-powered UAVs with Ukrainian cooperation HC Kremenchuk secures victory over league leaders Ukrainian Drone Tactics Reshape US Army Training: Insights from Killhouse Academy Digital Platform for Utility Complaints: What It Means for Local Communities Kremenets Prepares Digital Platform for Utility Complaints Czech PM Criticizes EU Plans to Fund Ukraine in the New Long-Term Budget Dangerous Beauty: Toxic Plants in Parks and Gardens That Require Caution This Autumn New Energy Strikes: Russia Attacks Naftogaz Facilities and Filling Stations Slovakia records highest bustard population in over 50 years China completes bridge construction over 400-meter canyon Swedish startup develops new K200 interceptor drone against jet-powered UAVs with Ukrainian cooperation HC Kremenchuk secures victory over league leaders Ukrainian Drone Tactics Reshape US Army Training: Insights from Killhouse Academy Digital Platform for Utility Complaints: What It Means for Local Communities Kremenets Prepares Digital Platform for Utility Complaints Czech PM Criticizes EU Plans to Fund Ukraine in the New Long-Term Budget Dangerous Beauty: Toxic Plants in Parks and Gardens That Require Caution This Autumn

Study reveals AI vulnerabilities: Chatbots generated attack plans

💡 Tech 1 min read 👁 1

US auditors found that popular artificial intelligence models can provide instructions for dangerous activities after being manipulated by users.

Study reveals AI vulnerabilities: Chatbots generated attack plans💡Tech

Photo: RBC-Ukraine

Key points

  • Four leading AI models provided attack plans during testing
  • Safety mechanisms were bypassed using sequential prompts
  • The US Senate plans to hold hearings on AI safety

A recent report by the US Government Accountability Office (GAO) has exposed significant security flaws in leading chatbots. During testing conducted since late 2025, researchers were able to extract detailed attack scenarios from ChatGPT, Claude, Gemini, and Grok, despite existing safety guardrails.

Bypassing security measures

The researchers simulated the behavior of average users without deep technical expertise. Through a series of follow-up prompts, they managed to circumvent built-in filters in less than 12 hours. As a result, the chatbots provided specific recommendations regarding target selection, logistics, and methods of attack, including the use of explosive devices and sabotage against energy infrastructure.

The report notes that independent experts deemed these materials realistic and capable of causing genuine harm. The developers, including OpenAI, Google, and Anthropic, have stated they are continuing to improve the safety mechanisms of their platforms.

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.