November 21, 2026 USD 44.93EUR 49.28PLN 11.06
Новини Подільськ — коротко й по суті
LIVE
Healthcare Infrastructure Recovery in Chernihiv Region: Progress and Partnerships What to do if your smartphone falls into water: 4 common mistakes EU payments to Russia for energy exceed financial aid to Ukraine Russian attack on Zaporizhzhia: kindergarten and residential building damaged Ukrainian delegation at COP31 blocks Russia's attempt to legitimize occupied territories via climate reports Tragedy in Kryvyi Rih: Paramedic and Child Killed in Ballistic Missile Strike Zelenskyy emphasizes vulnerability of Crimean Bridge as a legitimate target Night Attack on Zaporizhzhia: Kindergarten and Residential Building Damaged Massive energy strike: Rivne region faces power outages How to keep your home warm during winter: expert tips Healthcare Infrastructure Recovery in Chernihiv Region: Progress and Partnerships What to do if your smartphone falls into water: 4 common mistakes EU payments to Russia for energy exceed financial aid to Ukraine Russian attack on Zaporizhzhia: kindergarten and residential building damaged Ukrainian delegation at COP31 blocks Russia's attempt to legitimize occupied territories via climate reports Tragedy in Kryvyi Rih: Paramedic and Child Killed in Ballistic Missile Strike Zelenskyy emphasizes vulnerability of Crimean Bridge as a legitimate target Night Attack on Zaporizhzhia: Kindergarten and Residential Building Damaged Massive energy strike: Rivne region faces power outages How to keep your home warm during winter: expert tips

GAO report: Leading AI chatbots vulnerable to creating dangerous attack scenarios

💡 Tech 1 min read 👁 0

US auditors found that top artificial intelligence models can be manipulated into generating attack plans, despite existing safety guardrails.

GAO report: Leading AI chatbots vulnerable to creating dangerous attack scenarios💡Tech

Photo: RBC-Ukraine

Key points

  • Four leading AI models failed to block requests for attack plans
  • Safety bypass was achieved in under 12 hours through iterative questioning
  • Developers have pledged to enhance security following the GAO report

A comprehensive study by the US Government Accountability Office (GAO), conducted over nearly a year, has uncovered critical flaws in the safety mechanisms of leading AI chatbots, including ChatGPT, Claude, Gemini, and Grok. Experts demonstrated that even users without advanced technical expertise could bypass system restrictions by using iterative, probing prompts.

Methodology of bypassing safety

During the testing phase, researchers simulated the behavior of average users. While direct requests for attack plans were initially blocked, gradually refining the conversation allowed them to extract detailed instructions in under 12 hours. Consequently, the chatbots generated specific scenarios covering target selection, logistics, and methods for causing harm.

Security implications

According to the report, one platform provided a detailed scenario for using improvised explosive devices, while another outlined an attack on power grid infrastructure using drones. Senator Maggie Hassan emphasized that these findings necessitate immediate congressional action, as the accessibility of such tools poses a significant threat to public safety.

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.