November 19, 2026 USD 45.05EUR 49.53PLN 11.22
Новини Болград — коротко й по суті
LIVE
Latvia reaffirms commitment to diplomatic presence in Ukraine despite Russian strikes Donald Trump questions the rationale for further military support to Ukraine Donald Trump Criticizes European Aid to Ukraine US Discusses Security Model for Ukraine Based on NATO Article 5 Friedrich Merz outlines conditions for peace and potential sanctions relief against Russia Series of property crimes uncovered in Odesa region: suspects detained Europe to Participate in Future Negotiations with Russia Power outage schedules for November 20: current energy system status Life Despite the War: How an Izmail Seamstress Created Her Own Clothing Brand Deployment of European Ruta Block 1 Missiles in Operations Against Russian Infrastructure Latvia reaffirms commitment to diplomatic presence in Ukraine despite Russian strikes Donald Trump questions the rationale for further military support to Ukraine Donald Trump Criticizes European Aid to Ukraine US Discusses Security Model for Ukraine Based on NATO Article 5 Friedrich Merz outlines conditions for peace and potential sanctions relief against Russia Series of property crimes uncovered in Odesa region: suspects detained Europe to Participate in Future Negotiations with Russia Power outage schedules for November 20: current energy system status Life Despite the War: How an Izmail Seamstress Created Her Own Clothing Brand Deployment of European Ruta Block 1 Missiles in Operations Against Russian Infrastructure

Study reveals AI model vulnerability to manipulation in military scenarios

💡 Tech 1 min read 👁 1

Analysts have discovered that popular AI models can label civilian sites as military targets if prompted to treat the situation as a game.

Study reveals AI model vulnerability to manipulation in military scenarios💡Tech

Photo: RBC-Ukraine

Key points

  • AI models mistakenly identify civilian objects as military targets
  • Manipulation via 'game' scenarios bypasses safety filters
  • Developers have acknowledged the issue and are working on safeguards

A new study by Vals has demonstrated critical security gaps in leading AI models, such as Anthropic's Claude Opus 5.5, OpenAI, and Google. Experts tested the algorithms' ability to identify targets using satellite imagery and drone data covering regions in Ukraine and Iran.

Manipulation mechanisms and security risks

During testing, it was found that built-in safety filters are easily bypassed by simply changing the request context to a game scenario. Under these conditions, AI models begin to ignore ethical constraints and label civilians or surrendering soldiers as legitimate targets. Specifically, the Claude Opus 5.5 model identified an Iranian school as a military target in 72% of cases when the task was presented as a simulation.

Developers have already responded to the audit results. Representatives from Anthropic and OpenAI acknowledged the findings as concerning and stated they are working to strengthen safety barriers. While technical limitations, such as a lack of onboard computing power for drones, currently restrict the direct use of AI in weaponry, experts warn of the risks of using such tools to analyze combat footage and train targeting 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.