November 22, 2026 USD 44.93EUR 49.28PLN 11.06
Новини Ковель — коротко й по суті
LIVE
Kitchen Ergonomics: Why Massive Dining Tables Are Going Out of Style Explosions reported in Russia's Krasnodar following drone attack on Rosgvardia facility Donald Trump comments on the departure of press secretary Karoline Leavitt Archaeologists discover unique Roman infant burial in Great Britain Donald Trump publicly reaffirms friendly relations with Vladimir Putin SSU Alpha special forces neutralize over 4,300 invaders in one week Russia Admits Logistics Complications Following Crimean Bridge Attacks Finland seeks to recover 25 billion euros from Russia and Gazprom Frontline update: General Staff reports intense fighting and enemy shelling Demographic Challenges in Ukraine: Experts Assess the Impact of War on Population Kitchen Ergonomics: Why Massive Dining Tables Are Going Out of Style Explosions reported in Russia's Krasnodar following drone attack on Rosgvardia facility Donald Trump comments on the departure of press secretary Karoline Leavitt Archaeologists discover unique Roman infant burial in Great Britain Donald Trump publicly reaffirms friendly relations with Vladimir Putin SSU Alpha special forces neutralize over 4,300 invaders in one week Russia Admits Logistics Complications Following Crimean Bridge Attacks Finland seeks to recover 25 billion euros from Russia and Gazprom Frontline update: General Staff reports intense fighting and enemy shelling Demographic Challenges in Ukraine: Experts Assess the Impact of War on Population

Research reveals AI vulnerability to manipulation in military scenarios

💡 Tech 1 min read 👁 1

Analysts have discovered that modern AI models can flag civilian objects as military targets if the request is framed as a game.

Research reveals AI vulnerability to manipulation in military scenarios💡Tech

Photo: RBC-Ukraine

Key points

  • AI models bypass filters via game scenarios
  • Civilian objects are wrongly identified as targets
  • Developers are creating additional safety mechanisms

New research from Vals has demonstrated critical security gaps in systems developed by leading AI companies, including OpenAI, Google, and Anthropic. Experiments revealed that AI models can identify civilian objects and people as potential targets if a user manipulates the context settings.

Security bypass mechanisms

Researchers utilized real satellite and drone imagery covering parts of Ukraine and Iran. It was found that even simple framing of a task as a "game" allows the bypassing of internal safety filters. Specifically, the Claude Opus 5.5 model classified a civilian school near a military complex as a target in 72% of cases when the scenario was presented as a game.

Representatives from Anthropic and OpenAI acknowledged the audit results as concerning and stated they are working on strengthening safety barriers to prevent the use of AI in military applications.

While technical constraints, such as latency and limited onboard computing power for drones, currently hinder the integration of AI directly into weapons, experts warn of the risks of using these technologies for automated analysis of combat footage and training target acquisition systems.

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.