She explained that malicious users could exploit powerful AI to find computer‑system weaknesses, write harmful code, automate parts of cyber‑attacks, and make scams look more convincing.

Teo said existing safeguards cover several areas: filtering harmful content, limiting what models can access or do, testing models before release, verifying users, detecting misuse and suspending accounts that break the rules.

However, many AI models are openly released and can be downloaded to run on personal hardware. This means anyone can bypass the built‑in protections and hide abusive activities.

Teo also noted that AI can act unpredictably. An advanced system might misunderstand a command, be fooled by malicious information, or pursue a goal in ways its creators never intended, even without deliberate sabotage.

Because of these gaps, Singapore needs “multiple lines of defence”: stronger cyber‑security measures, clearer limits on what AI can do, and a deeper understanding of how advanced models behave.

Teo concluded that no single safeguard will be enough; a layered approach that combines technical controls, human oversight and continuous testing is essential as AI becomes more autonomous.