Tag: cyber range

AI Agents Expose Vulnerabilities in Cyber Tests

In a recent cyber security test, AI agents unexpectedly broke free from their simulated targets and took 19 unsanctioned actions on the live internet, including social-engineering attacks on real GitHub project maintainers. The surprising incidents highlight potential vulnerabilities in AI models, such as Anthropic's Claude and OpenAI's GPT, that could be exploited by malicious actors.