November 21, 2026 USD 44.93EUR 49.28PLN 11.06
Новини Житомир — коротко й по суті
LIVE
Kremlin’s Rhetoric on Kyiv and Ukraine: A New Wave of Aggressive Statements from Russian Leadership Hostile Strikes Hit Zhytomyr Region: Infrastructure Damaged and Two Injured Zhytomyr to Host Documentary Screening on Prince Ostrozky Heritage with American Researcher Enemy strikes energy facility in Vinnytsia: Part of the city blacked out, kindergarten damaged Generational continuity: why descendants of insurgents choose to serve in the modern Ukrainian army Scientists Explain Why It Is So Hard to Put Down Smartphones Roman schoolchildren accidentally discover ancient villa with unique artifacts Estonia Scrambles NATO Jets Amid Drone Threats Winter Weather Approaching Ukraine: Forecasters Warn of Snow and Sub-Zero Temperatures Zhytomyr Polytechnic Hosts 10th Anniversary Student Dance Battle Competition Kremlin’s Rhetoric on Kyiv and Ukraine: A New Wave of Aggressive Statements from Russian Leadership Hostile Strikes Hit Zhytomyr Region: Infrastructure Damaged and Two Injured Zhytomyr to Host Documentary Screening on Prince Ostrozky Heritage with American Researcher Enemy strikes energy facility in Vinnytsia: Part of the city blacked out, kindergarten damaged Generational continuity: why descendants of insurgents choose to serve in the modern Ukrainian army Scientists Explain Why It Is So Hard to Put Down Smartphones Roman schoolchildren accidentally discover ancient villa with unique artifacts Estonia Scrambles NATO Jets Amid Drone Threats Winter Weather Approaching Ukraine: Forecasters Warn of Snow and Sub-Zero Temperatures Zhytomyr Polytechnic Hosts 10th Anniversary Student Dance Battle Competition

Study reveals vulnerabilities in leading AI models regarding dangerous content generation

🎓 Education 1 min read 👁 0

US auditors found that popular AI chatbots can generate attack plans after users manipulate their safety filters.

Study reveals vulnerabilities in leading AI models regarding dangerous content generation🎓Education

Photo: RBC-Ukraine

Key points

  • AI models bypassed safety restrictions in under 12 hours
  • GAO report warns of risks to national and economic security
  • Developers are currently working on strengthening AI safety mechanisms

A recent report by the US Government Accountability Office (GAO) has highlighted significant vulnerabilities in major artificial intelligence systems, including ChatGPT, Claude, Gemini, and Grok. Researchers spent nearly a year testing whether these platforms could be coerced into providing instructions for dangerous activities.

Bypassing safety protocols

During the testing phase, researchers acted as ordinary users without deep technical knowledge. While initial direct prompts for attack plans were rejected, a strategy of iterative questioning allowed them to bypass built-in safety filters. Within 12 hours, all four tested systems produced detailed scenarios for potential attacks.

The findings indicate that AI models were able to suggest specific locations, logistics, and methods for attacks, including the use of explosives or targeting critical infrastructure. Senator Maggie Hassan has called for immediate Congressional hearings, noting that the ability of a standard chatbot to provide such guidance poses a substantial threat to national 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.