CSA: Take AI cyber threats to the boardroom

The highest levels of leadership at Singapore’s critical information infrastructure (CII) providers have been urged to urgently review their cyber defences, as rapid advances in artificial intelligence (AI) threaten to upend current risk management assumptions.

In an open letter to the boards and chief executives of CII owners, Singapore’s commissioner of cyber security and chief executive of the Cyber Security Agency, David Koh, warned that frontier AI has “materially shifted the cyber security baseline” in the past month.

These developments, he stressed, demand board-level and CEO attention and should not be left to IT departments.

The alert comes amid global scrutiny of new and highly capable AI models such as Anthropic’s Claude Mythos, which has shown unprecedented abilities to autonomously discover and exploit software vulnerabilities. Recognising its potent cyber capabilities, Anthropic restricted the model’s access to vetted defenders under an initiative dubbed Project Glasswing.

In his letter, Koh noted that Claude Mythos had already identified thousands of zero-day vulnerabilities. Shortly after it was released, the UK’s AI Security Institute reported that Mythos was the first model it tested that successfully completed a 32-step simulation of breaking into a corporate network, which would normally take an expert about 20 hours to do.

Meanwhile, OpenAI recently assessed its widely available GPT-5.5 model as having a “High” cyber security capability, just one step below “Critical” under its safety preparedness framework.

That means GPT-5.5 can conduct cyber operations against reasonably hardened targets or speed up discovery of vulnerabilities, while a model with a critical capability can go as far as developing zero-day exploits to compromise critical systems without human intervention.

“Frontier AI is accelerating at a rate where current assumptions in cyber risk management, on which your controls, measures and incident response plans were designed, may no longer be valid,” Koh wrote.

He noted that vulnerability discovery is also becoming faster and cheaper, social engineering is getting more personalised, and the window between a vulnerability’s disclosure and its exploitation by bad actors is rapidly narrowing.