Taiwan’s National Institute of Cybersecurity (NICS) will use agentic artificial intelligence to conduct black-box penetration tests during next year’s government cyber defense exercise, director Dennis Kung said at an Oct. 6 briefing, extending a ministry-wide effort to match AI-enabled attacks with AI-enabled defenses.
Four priorities Kung outlined
Kung told reporters in Taipei that his tenure will focus on four tracks: embedding agentic red teams in the annual Cybersecurity Operations and Defense Exercise, partnering with international AI vendors on models cleared for sensitive workloads, automating translation of global threat feeds into blocking rules for perimeter devices, and building a controlled range where agencies and companies can stress-test systems against high-intensity AI attacks.
How this fits the AI Basic Act
Taiwan’s Artificial Intelligence Basic Act took effect in January 2026, giving agencies until January 2028 to align sector rules with national principles. Cyber exercises are an early test bed because they generate logs, human oversight and after-action reviews that regulators can cite when drafting binding standards for government AI use.
The Financial Supervisory Commission and Ministry of Health and Welfare have published industry guidance; NICS work will likely inform cross-agency playbooks on autonomous security tools that can chain reconnaissance, exploitation and reporting steps without constant human clicks.
Vendor audits as backdrop
MODA’s Administration for Cyber Security consolidated 121 scheduled IT vendor audits this year into 16 joint reviews, finding recurring gaps in vulnerability scanning and firewall rule hygiene. Agentic red teams could continuously probe the same weaknesses between annual audits, Kung suggested.
Next year the joint audit program expands to 30 providers serving central and local governments, increasing the surface area NICS must cover.
Private-sector PSIRT push
Kung said NICS will help at least 30 domestic companies stand up product security incident response teams this year, mirroring expectations embedded in the European Union’s Cyber Resilience Act. Faster vulnerability disclosure reduces the window in which disguised Linux implants—such as those recently flagged on regional network gear—can persist undetected.
Risks agencies must manage
Autonomous attack tools can generate false positives or stray into production data if scopes are poorly defined. NICS officials said exercises will run in isolated environments with human approval gates before any exploit chain touches live citizen services.
For CIOs, the near-term takeaway is budget planning: red-team automation may shrink consulting bills but will require GPU capacity, log retention and legal review of model outputs. Kung pledged transparency reports after each exercise so legislators can judge whether agentic testing strengthens defenses or merely adds complexity.