The Most Impactful CVEs of 2025: A Year in Review

The Year in Vulnerabilities

2025 was defined by several trends: accelerating exploit timelines, continued targeting of edge devices, AI-related attack surfaces, and a record number of CVEs published. Here are the vulnerabilities that mattered most.

Edge Device Exploitation Continued

The pattern established in 2023–2024 intensified. VPNs, firewalls, and remote access gateways remained the #1 target for initial access by both ransomware groups and state-sponsored actors.

Ivanti Connect Secure / Policy Secure (Multiple CVEs)

Ivanti had a difficult 2025, with multiple rounds of vulnerabilities in their VPN and access control products. The cascade started with authentication bypass flaws that allowed unauthenticated remote code execution. Despite patches, many organizations struggled with remediation because the devices required careful upgrade procedures.

Lessons learned:

• Edge devices need the fastest patch cycles in your organization
• "Patch available" doesn't mean "patch applied", Ivanti's complex upgrade process delayed many organizations
• Network-level mitigations (restricting management interfaces) buy time while patching

Fortinet FortiGate

FortiGate firewalls saw critical pre-authentication vulnerabilities that were rapidly weaponized. Because these devices are internet-facing by design, exploitation required no special access, just knowledge of the target's IP address.

The speed from CVE publication to mass exploitation was measured in days, not weeks.

Supply Chain and Build System Attacks

CI/CD Pipeline Poisoning

2025 saw a marked increase in attacks targeting continuous integration systems. Compromised GitHub Actions, poisoned Docker images in popular registries, and malicious commits to widely-used build tools created ripple effects across thousands of downstream projects.

The XZ Utils backdoor (discovered in March 2024) set the template: patient, long-term compromise of open-source infrastructure by attackers playing the long game.

Key takeaway: Your build pipeline is part of your attack surface. Pin dependencies, verify checksums, and audit your CI/CD configurations regularly.

AI and LLM-Related Vulnerabilities

2025 was the first year where AI-specific vulnerabilities became a significant category:

Prompt Injection in Production Systems

As organizations deployed LLM-powered features, prompt injection emerged as a real attack vector, not just a research curiosity. Attackers found ways to manipulate AI agents into exfiltrating data, bypassing authorization checks, and executing unintended actions.

Model Serving Infrastructure

Vulnerabilities in model serving frameworks (TensorFlow Serving, Triton, vLLM) provided remote code execution paths. These tools are often deployed with high privileges and network access, making them attractive targets.

AI-Powered Phishing and Social Engineering

While not CVEs in the traditional sense, AI-generated phishing content reached a quality level that rendered many traditional detection methods ineffective. Voice cloning and video deepfakes were used in targeted attacks against executives.

Cloud and Kubernetes

Kubernetes Privilege Escalation

Multiple Kubernetes CVEs in 2025 allowed container escape or privilege escalation within clusters. The recurring pattern: a bug in how kubelet handles volume mounts or pod specifications allows a malicious container to gain host-level access.

Cloud Provider Service Vulnerabilities

Cross-tenant vulnerabilities in managed services, where one customer's actions could affect another customer's environment, remained a concern. These typically don't get CVE IDs but represent significant risk for cloud-native organizations.

The Trend Lines

Several trends accelerated through 2025:

• CVE volume keeps climbing, well over 40,000 CVEs assigned annually, up from ~29,000 in 2023
• KEV catalog growth, CISA is adding new confirmed-exploited vulnerabilities at a steady pace, with edge devices consistently overrepresented
• Time-to-exploitation is shrinking, industry reports consistently show the gap between disclosure and active exploitation getting shorter, with some high-profile CVEs weaponized within days
• Ransomware costs keep rising, IBM's 2024 report put average breach cost at $4.88M, and ransomware-specific incidents trend even higher

The direction is clear: more vulnerabilities, faster exploitation, and higher costs for organizations that fall behind.

What This Means for 2026

Prediction 1: Exploit timelines will continue shrinking. AI-assisted vulnerability research and exploit development will reduce the time from CVE publication to weaponized exploit. Organizations need faster patch processes, not just faster detection.

Prediction 2: Supply chain attacks will become more sophisticated. The success of dependency confusion, build system poisoning, and maintainer social engineering will attract more investment from advanced threat actors.

Prediction 3: AI attack surface will grow. As more organizations deploy LLM-powered features, AI-specific vulnerabilities (prompt injection, model poisoning, inference attacks) will become a standard part of the threat landscape.

Prediction 4: Edge devices remain the primary target. Until organizations consistently patch VPNs and firewalls within days of CVE disclosure, attackers will keep exploiting the patch gap.

How to Prepare

1. Shrink your patch window for internet-facing systems to 7 days or less
2. Generate and maintain SBOMs for supply chain visibility
3. Audit your CI/CD pipeline for trust assumptions and supply chain risks
4. Assess AI deployments for prompt injection and data exfiltration risks
5. Monitor the KEV catalog, it's the highest-signal threat intelligence feed available

The organizations that navigated 2025 best were those that prioritized speed over completeness in their vulnerability response. You don't need to fix everything. You need to fix the right things fast.