November 18, 2026 USD 45.20EUR 49.57PLN 11.24
Новини Славута — коротко й по суті
LIVE
Cooperation with Portugal: New Steps to Strengthen Air Defense and Energy Resilience Experts determine the optimal frequency for washing jeans to maintain their appearance Funding of the ROC by Russian businessmen should become grounds for international sanctions Energy Infrastructure Struck in Odesa Region: Impact and Situation Update Civilian evacuation from Kramatorsk: 11 more residents rescued Appointment of sanctioned Russian official to IAEA sparks outrage in Ukraine Armed Forces of Ukraine adopt NATO standards: sergeants gain authority for independent decision-making Deaths of despair: new study reveals alarming trends in Europe Kateryna Repiakhova responds to criticism regarding age gap with Viktor Pavlik Strikes on Russian Refineries: Oil Processing Volumes in Russia Significantly Reduced Cooperation with Portugal: New Steps to Strengthen Air Defense and Energy Resilience Experts determine the optimal frequency for washing jeans to maintain their appearance Funding of the ROC by Russian businessmen should become grounds for international sanctions Energy Infrastructure Struck in Odesa Region: Impact and Situation Update Civilian evacuation from Kramatorsk: 11 more residents rescued Appointment of sanctioned Russian official to IAEA sparks outrage in Ukraine Armed Forces of Ukraine adopt NATO standards: sergeants gain authority for independent decision-making Deaths of despair: new study reveals alarming trends in Europe Kateryna Repiakhova responds to criticism regarding age gap with Viktor Pavlik Strikes on Russian Refineries: Oil Processing Volumes in Russia Significantly Reduced

AI Vulnerability: How Manipulations Cause Neural Networks to Target Civilian Objects

💡 Tech 1 min read 👁 0

Research has uncovered critical security flaws in leading AI models that allow users to bypass restrictions on military-related queries.

AI Vulnerability: How Manipulations Cause Neural Networks to Target Civilian Objects💡Tech

Photo: RBC-Ukraine

Key points

  • AI models are vulnerable to manipulation via game scenarios
  • Neural networks mistakenly identify civilian objects as military targets
  • Developers promise to strengthen defenses against military misuse

A new independent study by Vals analysts has demonstrated significant gaps in the security architecture of leading AI models, including those from OpenAI, Google, and Anthropic. Experiments revealed that neural networks can identify civilian objects and people as military targets if the user employs manipulative techniques.

Manipulation mechanisms and consequences

Researchers found that the models' safety filters are easily bypassed if a request is framed as part of a "game scenario." In tests using real satellite imagery, the Claude Opus 5.5 model identified civilian objects, such as schools, as potential targets in over 70% of cases.

Representatives from the developers, including Anthropic and OpenAI, have acknowledged these results as concerning. They announced plans to strengthen safety protocols and implement additional barriers to prevent the use of AI for military purposes or drone data analysis.

Source RBC-Ukraine — rbc.ua/ukr/news/dostatno-nazvati-boyovi-diyi-groyu-claude-179501202… ↗

Contact the newsroom

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