Cyber defense in the era of frontier AI: Insights from Mythos and GPT 5.5 Cyber

Zscaler evaluates frontier AI models (Mythos and GPT-5.5 Cyber) across black-box, artifact/code, and gray/white-box test harnesses, finding that multi-step reasoning enables the models to chain vulnerabilities into actionable attack paths, surface higher-fidelity high-severity findings, and accelerate detection, triage, and remediation. The report recommends defensive measures including Zero Trust, hiding applications, AI guardrails, deception, and prioritized exposure management to mitigate risks posed by attackers using similar AI capabilities.