Zscaler Report Finds Enterprises Unprepared for Frontier AI Cyberattacks
Summary
Zscaler says hundreds of assessed enterprises averaged 37/100 for frontier AI readiness, citing exposed VPNs, weak AI identity controls, and broad access rules. It recommends a 90-day push on Zero Trust controls, patching, traffic inspection, and deception.
Key points
- The hypothetical breach scenario describes attackers exploiting a pre-authentication VPN flaw, stealing credentials, and using an identity trust chain to reach internal AI systems and proprietary data.
- Across hundreds of enterprises assessed in 2026, the average Frontier AI Readiness Score was 37/100; the AI governance pillar averaged 2.1/20.
- Nine in ten assessed organizations had exposed VPN appliances, and more than one-third left CISA Known Exploited Vulnerabilities active on internet-facing assets.
- All assessed organizations lacked identity-based authentication for AI systems; enterprises inspected an average of about 36% of encrypted traffic.
- Broad segmentation configurations were present at 64% of organizations; one case involved 195 wildcard domains exposing more than 108,000 applications.
- Zscaler recommends enforcing dormant security controls, shielding VPNs and patching known exploited flaws, requiring identity checks for AI endpoints, increasing TLS inspection, and replacing wildcard rules with named domains and decoys.
Article Details
- Publisher
- Zscaler
- Report Period
- 2026
- Scope
- Frontier AI readiness assessments of enterprises, using isolated, controlled simulation environments rather than live enterprise data.
- Sample Size
- Hundreds of enterprises; exact number not disclosed.
- Key Statistics
- Average Frontier AI Readiness Score: 37 out of 100 across assessed enterprises.
- Average AI governance pillar score: 2.1 out of 20.
- 9 out of 10 assessed organizations had exposed VPN appliances.
- More than one-third of enterprises had active CISA Known Exploited Vulnerabilities on internet-facing assets.
- All assessed organizations lacked identity-based authentication for their AI systems.
- Recommendations
- Move dormant controls, including IPS, malicious IP blocking, and domain fronting protection, from monitor-only to enforcement.
- Put exposed VPN appliances behind brokered Zero Trust access and promptly patch known exploited vulnerabilities on public-facing assets; hide assets that cannot be patched.
- Require enterprise identity verification for AI endpoints and copilots, and raise SSL/TLS inspection to at least 70% by default.
- Replace wildcard domain rules with named FQDNs and deploy active decoys near critical applications to detect and revoke compromised sessions.