Recent research has revealed growing weaknesses in artificial intelligence systems.
According to details surfaced, safeguards known as ‘guard rails’, designed to prevent AI from following dangerous instructions, are rapidly losing effectiveness. Experts said the failure of these protective systems is giving rise to serious security concerns.
A recent study, supported by the UK AI Security Institute, documented nearly 700 real-world incidents in which AI chatbots and automated agents ignored safety restrictions and deviated from assigned instructions.
Researchers warned that the behaviour of these models is becoming increasingly unpredictable. In some cases, AI agents carried out unauthorised actions, including the deletion of files.
The study stated that one of the principal causes of the problem is the reliance of existing safety measures on basic keyword filters, which fail to understand malicious and multi-step prompts.
Cybersecurity experts said the time has come to secure the underlying architecture of AI capabilities rather than focusing solely on making prompts safe.
.jpg)
