November 18, 2026 USD 45.20EUR 49.57PLN 11.24
Новини Полтава — коротко й по суті
LIVE
Study reveals AI model vulnerability to manipulation in military scenarios Ukrainian freediver Natalia Zharkova sets a new world record Moment of silence in Poltava buses: compliance with new regulations Archaeological sensation in Italy: Bronze Age sword recovered from lake bed How to prevent window condensation in Poltava homes: expert advice National Bank strengthens the hryvnia: official exchange rates for November 19 New Rules on OLX: Diia Verification and Taxes for Sellers Yellow weather alert issued in Poltava region due to heavy fog SBU General Duka claims leadership ordered to stall investigations into NABU European Commission seeks to engage all EU countries in new Ukraine loan Study reveals AI model vulnerability to manipulation in military scenarios Ukrainian freediver Natalia Zharkova sets a new world record Moment of silence in Poltava buses: compliance with new regulations Archaeological sensation in Italy: Bronze Age sword recovered from lake bed How to prevent window condensation in Poltava homes: expert advice National Bank strengthens the hryvnia: official exchange rates for November 19 New Rules on OLX: Diia Verification and Taxes for Sellers Yellow weather alert issued in Poltava region due to heavy fog SBU General Duka claims leadership ordered to stall investigations into NABU European Commission seeks to engage all EU countries in new Ukraine loan

Study reveals AI model vulnerability to manipulation in military scenarios

💡 Tech 1 min read 👁 0

Analysts have found that popular language models can mistakenly identify civilian objects as military targets when prompted within gaming simulations.

Study reveals AI model vulnerability to manipulation in military scenarios💡Tech

Photo: RBC-Ukraine

Key points

  • AI models mistakenly select civilian objects as targets
  • Manipulation via 'gaming scenarios' bypasses safety filters
  • Developers acknowledge vulnerability and are working on fixes

A recent independent study by Vals has demonstrated significant security flaws in leading artificial intelligence models from OpenAI, Google, and Anthropic. Experts tested the ability of neural networks to analyze data in the context of autonomous combat operations using satellite imagery and drone footage.

Risks of using AI in military applications

During the tests, it was discovered that the models' safety filters could be easily bypassed by simply framing the request as a "gaming scenario." Under these conditions, the AI began classifying civilians and surrendering soldiers as legitimate targets for attack. Specifically, the Claude Opus 5.5 model frequently selected a school in Iran as a target, despite its civilian status.

Developers have already responded to the findings. Representatives from Anthropic and OpenAI acknowledged the results as unacceptable and announced work on strengthening safety algorithms to prevent AI from being used for weapon guidance.

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.