Existing safeguards by AI companies insufficient as models grow more capable: Josephine Teo

Minister Josephine Teo of Singapore highlighted that existing safeguards by AI companies are insufficient to address risks from increasingly capable models. While content filtering, user verification, and pre-release testing exist, malicious users can bypass them by using openly available AI models, she stated in a Facebook post.

Teo warned that such systems could be used to exploit computer systems, write malicious code, automate cyberattacks, and craft more convincing scams. Although safeguards are crucial, malicious actors can still leverage openly available models to circumvent them. Additionally, AI systems may misunderstand instructions, be deceived by malicious information, or perform unintended actions due to over-authority.

The government must adopt multiple defense layers, enhance cyber defenses, improve attack detection and recovery, and carefully balance convenience and security. The Cyber Security Agency of Singapore has issued guidance on patching vulnerabilities, strong authentication, and tightening access controls. AI can also be used defensively to identify vulnerabilities.

Even without advanced AI access, organizations can leverage available AI tools for cyber defense. When AI models have access to data, tools, and processes, safeguards become even more critical. AI agents can cause harm unintentionally by misunderstanding instructions, pursuing unintended goals, or being tricked by malicious information.

The greater the potential impact, the stronger the safeguards and human oversight should be. Singapore's AI Safety Institute is building technical capabilities to evaluate advanced AI systems, and the country supports international collaboration on testing and evaluation. Teo emphasized that a multi-layered approach, including stronger cyber defenses, clear limits on AI actions, appropriate human oversight, and testing capabilities, is essential to build trust in AI serving the public good.