November 18, 2026 USD 45.20EUR 49.57PLN 11.24
Новини Біла Церква — коротко й по суті
LIVE
An Amber Throne for Donald Trump: A Secret Gift from the Polish President Media reports suggest potential Trump-Putin meeting at China summit Ukrainian blogger Oleksandr Slobozhenko detained in Italy: extradition process underway 20-year-old arrested in Poland for shooting threats: 'Prank' leads to criminal charges International operation: Ukrainian authorities and FBI dismantle ransomware hacking group AI Vulnerability: How Manipulations Cause Neural Networks to Target Civilian Objects Ukrainian blogger Oleksandr Slobozhenko detained in Italy: extradition process underway National Bank of Ukraine strengthens the hryvnia: official exchange rates for November 19 After eight months of struggle: 12-year-old from Bila Tserkva returns home after life-saving treatment at Okhmatdyt SBU General claims leadership interfered with investigations into NABU An Amber Throne for Donald Trump: A Secret Gift from the Polish President Media reports suggest potential Trump-Putin meeting at China summit Ukrainian blogger Oleksandr Slobozhenko detained in Italy: extradition process underway 20-year-old arrested in Poland for shooting threats: 'Prank' leads to criminal charges International operation: Ukrainian authorities and FBI dismantle ransomware hacking group AI Vulnerability: How Manipulations Cause Neural Networks to Target Civilian Objects Ukrainian blogger Oleksandr Slobozhenko detained in Italy: extradition process underway National Bank of Ukraine strengthens the hryvnia: official exchange rates for November 19 After eight months of struggle: 12-year-old from Bila Tserkva returns home after life-saving treatment at Okhmatdyt SBU General claims leadership interfered with investigations into NABU

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

💡 Tech 2 min read 👁 0

Research by Vals has revealed that leading AI models can be easily deceived by framing military scenarios as games, leading them to identify civilian objects as military targets.

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

Photo: RBC-Ukraine

Key points

  • AI models easily bypass safety filters through game scenarios
  • Neural networks mistakenly identify civilian objects as military targets
  • Developers have acknowledged the issue and are working on new barriers

Dangerous Security Gaps in AI

A new study by analysts at Vals has demonstrated critical vulnerabilities in the algorithms of popular AI models from OpenAI, Google, and Anthropic. Experiments showed that neural networks, which have built-in safeguards against assisting in military operations, can be easily bypassed through a simple manipulation—framing the prompt as a "game."

During the tests, the models analyzed real satellite imagery and drone footage. As a result, the AI regularly flagged civilians and surrendering soldiers as legitimate targets. A particularly striking example was the Claude Opus 5.5 model, which selected a civilian school in Iran as a target in 72% of cases when the scenario was presented as a game.

Developer Response and Risks

Representatives of the developer companies, including Anthropic and OpenAI, have acknowledged these results as unacceptable and announced the strengthening of security protocols. However, experts warn that malicious actors could use publicly available models to process massive amounts of battlefield video data, posing serious ethical and security challenges.

While technical limitations, such as the insufficient processing power of onboard drone computers, currently hinder the autonomous application of AI, the rapid development of these technologies may change this situation in the near future.

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.