AI cybersecurity threats begin with exposed systems
The 2023 vulnerability advisory is not evidence of AI-driven exploitation. It documents something more familiar and still urgent: attackers repeatedly exploited weaknesses in enterprise products, including flaws used as zero-days. That distinction matters. AI services do not change what this report says; they do give defenders another reason to secure the infrastructure on which important services depend. NSA annual report announcement
The advisory was produced by U.S. agencies and Five Eyes partners. Its 15 vulnerabilities are those the agencies identified as routinely exploited in 2023. The list is useful for triage, not exhaustive: the advisory also discusses 32 other vulnerabilities often exploited that year. An absent CVE is not a declaration of safety. Canadian Centre for Cyber Security joint advisory; BleepingComputer report
What the 2023 exploitation data says
Eleven of the 15 vulnerabilities were initially exploited as zero-days, BleepingComputer reports, compared with less than half of the top exploited vulnerabilities in 2022. The Canadian advisory makes the broader comparison: malicious actors exploited more zero-days to compromise enterprise networks in 2023 than in 2022, enabling operations against higher-priority targets. Those are related but distinct observations—the first concerns the top-15 set; the second describes enterprise-network exploitation more broadly. BleepingComputer report; Canadian Centre for Cyber Security joint advisory
A zero-day is exploited before a patch is available. That leaves defenders less room to get ahead of the activity. Once a fix is released, exposure does not vanish on its own: teams still have to identify affected assets and remediate them. The agencies also noted that many entries were in the top 15 for the first time, a reminder that historical priority lists cannot replace current visibility. BleepingComputer report
The reported top 15 are:
| CVE | Vendor and product | Vulnerability type |
|---|---|---|
| CVE-2023-3519 | Citrix NetScaler ADC/Gateway | Code injection |
| CVE-2023-4966 | Citrix NetScaler ADC/Gateway | Buffer overflow |
| CVE-2023-20198 | Cisco IOS XE Web UI | Privilege escalation |
| CVE-2023-20273 | Cisco IOS XE | Web UI command injection |
| CVE-2023-27997 | Fortinet FortiOS/FortiProxy SSL-VPN | Heap-based buffer overflow |
| CVE-2023-34362 | Progress MOVEit Transfer | SQL injection |
| CVE-2023-22515 | Atlassian Confluence Data Center/Server | Broken access control |
| CVE-2021-44228 (Log4Shell) | Apache Log4j2 | Remote code execution |
| CVE-2023-2868 | Barracuda Networks ESG appliance | Improper input validation |
| CVE-2022-47966 | Zoho ManageEngine products | Remote code execution |
| CVE-2023-27350 | PaperCut MF/NG | Improper access control |
| CVE-2020-1472 | Microsoft Netlogon | Privilege escalation |
| CVE-2023-42793 | JetBrains TeamCity | Authentication bypass |
| CVE-2023-23397 | Microsoft Outlook | Privilege escalation |
| CVE-2023-49103 | ownCloud graphapi | Information disclosure |
The complete CVE, product, and type entries above follow the published list. Check the linked advisory and the vendor’s current guidance when assessing a particular deployment; the table is not a substitute for determining whether a system is affected. BleepingComputer report
Use the list to prioritize, not to draw a boundary
Start with inventory. Find whether your organization runs the affected products, identify versions and exposure, and establish who owns each system. Put internet-facing and business-critical assets near the front of the queue, then confirm that patches or mitigations were actually applied. The advisory’s list can sharpen that review, but it cannot tell you that other vulnerabilities are harmless. Canadian Centre for Cyber Security joint advisory
The agencies recommend timely patching and centralized patch management. They also recommend defensive tools including endpoint detection and response, web application firewalls, and network protocol analyzers. These are complementary controls: monitoring helps identify suspicious activity and remaining exposure, while patching removes known weaknesses where a fix can be applied. Keep exceptions visible, assign an owner, and revisit them rather than allowing a temporary delay to become permanent. NSA annual report announcement; BleepingComputer report
The qualified connection to AI and agent security
The reporting does not attribute these vulnerabilities or their exploitation to AI, an AI agent, or agentic systems. It would be a mistake to turn an exploited-CVE report into a claim about autonomous attacks. The defensible connection is narrower: AI-enabled services still rely on ordinary software and infrastructure, and those supporting systems need sound vulnerability management.
For teams securing AI services, include the surrounding environment in the same inventory and patch process used for other business-critical technology. That is a practical defensive implication, not a finding made by the 2023 advisory. The report gives no basis for claims about AI-specific attack methods, model behavior, or agent incidents, so those claims do not belong here.
Ask vendors about secure defaults
The NSA announcement urges vendors, designers, and developers to prioritize secure-by-default configurations and says published CVEs should include appropriate Common Weakness Enumerations. It also recommends that buyers ask software providers about secure-by-design programs. These steps complement operational patching: providers have responsibilities in product design, while customers still need to know what they run and address exposure in their own environments. NSA annual report announcement
The useful takeaway for 2026 is not that this 2023 set predicts every current threat. It is evidence of what agencies observed then: routine exploitation, substantial zero-day use, and a need for prompt remediation and better visibility. Use the list to ask focused questions, not to stop looking beyond it. For AI services and everything else, that is the less glamorous—and more reliable—security practice.
ai cybersecurity threats begin with exposed systems