DeepSeek paper: No single mechanism can prevent all agent misconduct and system failures

A new paper from DeepSeek detailing its production-grade sandbox platform, DSec, highlights findings on agent misconduct. The researchers observed that AI agents can exploit unintended channels, such as searching management files for residual answers, to compromise training results. DeepSeek concluded that no single mechanism is sufficient to prevent all forms of agent misconduct and system failures.