November 19, 2026 USD 45.05EUR 49.53PLN 11.22
Новини Покров — коротко й по суті
LIVE
Zelenskyy calls for foreign military bases in Ukraine after the war President Zelensky: Russia attempts to compensate for battlefield losses with civilian attacks Irregular sleep increases heart disease risk, study finds Banana Ripeness: How to Choose the Right Fruit for Your Health Needs Fuel prices in Ukraine to drop: forecasts for Pokrov drivers Secrets of Longevity: Why Natural Activity Outperforms Gym Workouts Donald Tusk comments on dialogue with Donald Trump regarding negotiations with Russia Scientists determine the exact date of the fall of ancient Babylon Emmanuel Macron: Diplomatic channels with Russia remain open Latvian diplomats continue work in Ukraine despite Russian strikes Zelenskyy calls for foreign military bases in Ukraine after the war President Zelensky: Russia attempts to compensate for battlefield losses with civilian attacks Irregular sleep increases heart disease risk, study finds Banana Ripeness: How to Choose the Right Fruit for Your Health Needs Fuel prices in Ukraine to drop: forecasts for Pokrov drivers Secrets of Longevity: Why Natural Activity Outperforms Gym Workouts Donald Tusk comments on dialogue with Donald Trump regarding negotiations with Russia Scientists determine the exact date of the fall of ancient Babylon Emmanuel Macron: Diplomatic channels with Russia remain open Latvian diplomats continue work in Ukraine despite Russian strikes

AI Vulnerability: Anthropic and OpenAI models may misidentify civilian targets as military ones

💡 Tech 1 min read 👁 3

A new study by Vals analysts has revealed critical flaws in AI systems that allow them to be manipulated into targeting civilian infrastructure.

AI Vulnerability: Anthropic and OpenAI models may misidentify civilian targets as military ones💡Tech

Photo: RBC-Ukraine

Key points

  • AI models are vulnerable to manipulation via game-like scenarios
  • Models misidentify civilian targets as military ones
  • Developers have acknowledged the issue and are working on safeguards

Analytics firm Vals has released a report demonstrating dangerous security gaps in advanced artificial intelligence models, including those from OpenAI and Anthropic. Experiments revealed that AI systems can identify civilians and social facilities as potential military targets if a user employs specific manipulative techniques.

Manipulation mechanisms and developer response

Researchers conducted tests by combining simulations with actual satellite imagery of locations in Ukraine and Iran. They discovered that simply framing a scenario as a "game" is enough to bypass built-in safety filters. For instance, the Claude Opus 5.5 model classified a school near a military complex as a target in the majority of tests.

Developers have already responded to the findings. Anthropic representative Eduardo Maia Silva called the results unacceptable, while OpenAI stated they are actively strengthening barriers to prevent the use of their models for military purposes. Currently, technical limitations such as latency and insufficient onboard computing power prevent the direct integration of these models into combat drones.

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.