November 18, 2026 USD 45.20EUR 49.57PLN 11.24
Новини Кропивницький — коротко й по суті
LIVE
AI Vulnerability: How Manipulations Force Neural Networks to Target Civilian Objects Shakhtar Donetsk loses Gleyker Mendoza until the end of 2026 due to injury Energy experts predict cyclical power outages for the next two months Euro 2028: Ukraine awaits draw, Russia excluded from tournament SBU General Serhiy Duka accuses anti-corruption agencies of political pressure Sports Isolation of Russia: How European Countries Bypass International Federation Rulings Shevchenkivskyi District Court orders Prosecutor General's Office to process Kolomoisky's complaint German Foreign Minister comments on upcoming talks between US, Chinese, and Russian leaders National Bank strengthens hryvnia: official exchange rates for November 19 New Rules on OLX: Identity Verification via 'Diia' and Tax Changes for Sellers AI Vulnerability: How Manipulations Force Neural Networks to Target Civilian Objects Shakhtar Donetsk loses Gleyker Mendoza until the end of 2026 due to injury Energy experts predict cyclical power outages for the next two months Euro 2028: Ukraine awaits draw, Russia excluded from tournament SBU General Serhiy Duka accuses anti-corruption agencies of political pressure Sports Isolation of Russia: How European Countries Bypass International Federation Rulings Shevchenkivskyi District Court orders Prosecutor General's Office to process Kolomoisky's complaint German Foreign Minister comments on upcoming talks between US, Chinese, and Russian leaders National Bank strengthens hryvnia: official exchange rates for November 19 New Rules on OLX: Identity Verification via 'Diia' and Tax Changes for Sellers

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

💡 Tech 2 min read 👁 0

Researchers have discovered that popular AI models can be manipulated into ignoring ethical constraints if requests are framed as game scenarios.

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

Photo: RBC-Ukraine

Key points

  • AI models bypass ethical filters when prompted as games
  • Civilian objects are erroneously classified as military targets
  • Developers have acknowledged the issue and are working on security patches

Risks of AI in Military Applications

A new study by Vals analysts has exposed a critical vulnerability in modern AI models, including Claude Opus 5.5, OpenAI, and Google. Experiments revealed that these neural networks are capable of identifying civilian infrastructure as military targets if the user manipulates the context of the query.

During tests using satellite imagery of real locations in Ukraine and Iran, the models consistently classified civilians and prisoners of war as legitimate targets. It is particularly concerning that the systems' safety filters are easily bypassed by simply framing the task as a "game."

The Challenge of Ethical Barriers

According to the researchers, the Claude Opus 5.5 model selected a real school in Iran as a target in over 70% of cases, regardless of whether the prompt was presented as a serious scenario or entertainment. This highlights the inefficiency of current safeguards against technology misuse.

Representatives from developers, including Anthropic and OpenAI, acknowledged these results as unacceptable and stated they are working to strengthen security measures. Currently, technical limitations, such as insufficient onboard computing power for drones, hinder the direct integration of AI into weapon systems, yet the threat of hackers using these tools to analyze video feeds remains a significant concern.

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.