November 18, 2026 USD 45.20EUR 49.57PLN 11.24
Новини Дніпро — коротко й по суті
LIVE
Polish police arrest 20-year-old for online shooting threats Fines for smoking in Dnipro educational institutions: what you need to know AI vulnerability: Models can misidentify civilian targets through simple manipulation Businesses in Dnipro can acquire solar power plants without initial investment Fines for Smoking on School Grounds in Dnipro: What Parents and Students Need to Know National Bank of Ukraine strengthens the hryvnia: exchange rates for November 19 Water supply in Dnipro during blackouts: how utilities pump water to upper floors Youth Council of Kamianske hosts board game evening Ukraine's Defense Industry: Laser Weapons and Air Defense Challenges OLX to Implement Mandatory Verification via 'Diia' for Sellers Polish police arrest 20-year-old for online shooting threats Fines for smoking in Dnipro educational institutions: what you need to know AI vulnerability: Models can misidentify civilian targets through simple manipulation Businesses in Dnipro can acquire solar power plants without initial investment Fines for Smoking on School Grounds in Dnipro: What Parents and Students Need to Know National Bank of Ukraine strengthens the hryvnia: exchange rates for November 19 Water supply in Dnipro during blackouts: how utilities pump water to upper floors Youth Council of Kamianske hosts board game evening Ukraine's Defense Industry: Laser Weapons and Air Defense Challenges OLX to Implement Mandatory Verification via 'Diia' for Sellers

AI vulnerability: Models can misidentify civilian targets through simple manipulation

💡 Tech 1 min read 👁 0

Security researchers have found that leading AI models can be manipulated into identifying civilian infrastructure as military targets by framing tasks as games.

AI vulnerability: Models can misidentify civilian targets through simple manipulation💡Tech

Photo: RBC-Ukraine

Key points

  • AI models bypassed safety filters using game context
  • Civilian schools misidentified as military targets
  • Developers are working on enhanced safety barriers

A recent audit by Vals analysts has exposed significant security flaws in artificial intelligence models developed by major firms like OpenAI, Google, and Anthropic. The study demonstrated that these models are susceptible to manipulation, allowing users to bypass built-in safety filters simply by framing scenarios as a game.

Risks of autonomous targeting

During testing, the Claude Opus 5.5 model identified civilian objects, including schools, as potential targets in 72% of cases when the task was presented as a game. This highlights a dangerous trend where AI systems, lacking true situational awareness, can be misled into making lethal decisions based on flawed input.

Tech companies have acknowledged these findings, with Anthropic and OpenAI pledging to implement stricter safeguards to prevent the misuse of their models in military or drone-based applications. While technical constraints currently limit the deployment of such models on autonomous drones, experts warn that public access to these tools could facilitate the training of weapon systems, posing a significant risk to global security.

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.