November 19, 2026 USD 45.05EUR 49.53PLN 11.22
Новини Нетішин — коротко й по суті
LIVE
Another Enemy Attack on Kyiv: Residential Fires and One Injured Reported Working during blackouts: how Ukrainian businesses are adapting offices and remote formats Moscow Invites EU Representatives to Talks on Ukraine: Details of Putin's Statement Kremlin Statements Amid International Talks: Russian Leadership's Stance on Ukraine Pensioner and 600 pigeons evacuated from frontline Kramatorsk Occupants announce launch of boats across Kerch Strait after new bridge strikes From Professor to Disillusioned Collaborator: Yevhenia Bilchenko Detained in St. Petersburg Massive Afforestation in China: Scientists Warn of Hidden Environmental Risks UN International Court of Justice First Finds a State Responsible for Genocide Panic in Crimea Following Kerch Bridge Damage: Food Shortages and Fuel Limits Another Enemy Attack on Kyiv: Residential Fires and One Injured Reported Working during blackouts: how Ukrainian businesses are adapting offices and remote formats Moscow Invites EU Representatives to Talks on Ukraine: Details of Putin's Statement Kremlin Statements Amid International Talks: Russian Leadership's Stance on Ukraine Pensioner and 600 pigeons evacuated from frontline Kramatorsk Occupants announce launch of boats across Kerch Strait after new bridge strikes From Professor to Disillusioned Collaborator: Yevhenia Bilchenko Detained in St. Petersburg Massive Afforestation in China: Scientists Warn of Hidden Environmental Risks UN International Court of Justice First Finds a State Responsible for Genocide Panic in Crimea Following Kerch Bridge Damage: Food Shortages and Fuel Limits

AI Vulnerability: How Manipulation Forces Neural Networks to Target Civilian Objects

💡 Tech 1 min read 👁 0

Research has uncovered a critical vulnerability in popular AI models, which can identify civilian objects as military targets when subjected to manipulative prompts.

AI Vulnerability: How Manipulation Forces Neural Networks to Target Civilian Objects💡Tech

Photo: RBC-Ukraine

Key points

  • AI models mistakenly target civilian objects
  • Manipulation via 'game' context bypasses safety filters
  • Developers are creating new security barriers

An independent study by Vals analysts has revealed dangerous gaps in the security systems of leading artificial intelligence models, including those from OpenAI, Google, and Anthropic. Experiments demonstrated that neural networks, despite having safety filters, are easily manipulated if a user frames the request as part of a "game."

Risks of AI in Military Scenarios

During tests using real drone and satellite imagery, the Claude Opus 5.5 model frequently classified civilian objects, such as schools, as potential targets for attack. Notably, the frequency of these errors increased when the scenario was presented as a game.

Developers have acknowledged the audit results as alarming. Representatives from Anthropic and OpenAI stated they are working to strengthen safety mechanisms to prevent their technologies from being used for weapon guidance. Currently, technical limitations, such as data latency, are hindering the integration of AI directly into autonomous combat drones.

Experts warn that the accessibility of these tools allows malicious actors to analyze vast amounts of video data to automate target acquisition.

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.