As reported by The Hacker News, CISA has added four vulnerabilities to its Known Exploited Vulnerabilities (KEV) catalog, all showing evidence of active exploitation in the wild. The additions span an unusually broad attack surface — from legacy Java SOAP libraries to AI/LLM orchestration tooling — and underscore how threat actors are opportunistically targeting both aging infrastructure and emerging tech stacks simultaneously.
Vulnerability Summary
| CVE | CVSS | Product | Type | Exploitation |
|---|---|---|---|---|
| CVE-2026-90436 | 9.8 (Critical) | Apache Axis | Unsafe Reflection → RCE | Active — cryptomining via Confluence XML-RPC/SOAP |
| CVE-2026-90817 | 9.8 (Critical) | Vanderbilt REDCap | Code Injection → RCE | Active — ransomware campaigns |
| CVE-2026-40217 | 8.8 (High) | BerriAI LiteLLM | Unprotected Alternate Channel → RCE | Active — LLMjacking |
| CVE-2024-10441 | 9.8 (Critical) | Synology DSM | Improper Encoding/escaping → RCE | Active |
Why This Matters
The most striking aspect of this KEV batch is the diversity of attack vectors and downstream payloads. We're seeing a convergence where the same opportunistic playbook — scan for exposed services, exploit unauthenticated RCE, establish persistence — now targets AI infrastructure (LiteLLM) alongside traditional enterprise platforms. The LLMjacking angle on CVE-2026-40217 is particularly noteworthy: attackers aren't just after data, they're after compute. LLM API proxies are becoming high-value targets for criminals who want to resell access to frontier models or run unauthorized inference workloads at someone else's expense.
The Apache Axis exploitation chain detailed by Rapid7 is equally instructive. The vulnerability isn't in Atlassian Confluence directly — it's in a bundled legacy library that Confluence uses for XML-RPC and SOAP APIs. Attackers exploited CVE-2026-90436 to execute shell commands, fetch payloads from Cloudflare Tunnel and Pinggy endpoints, and within three minutes were authenticating to MySQL using Confluence's own database credentials. That timeline — three minutes from initial execution to credential access — is a stark reminder of how thin the window is between initial compromise and deep traversal.
The Confluence attack path is especially insidious because it targets a feature (Remote API via XML-RPC/SOAP) that was disabled by default starting in Confluence 3.0 and fully removed in Confluence 10.0. Organizations running older but still-supported Confluence versions with this feature enabled are the exposed population.
Who Is at Risk
/guardrails/test_custom_code endpoint without authentication.Shield53 Recommendations
Immediate Actions
- Patch all four affected products immediately. Check vendor advisories for Synology DSM, BerriAI LiteLLM, Apache Axis, and Vanderbilt REDCap for fixed versions.
- Disable Confluence Remote API (XML-RPC & SOAP) if still enabled. This setting should be off in any modern deployment. Audit Confluence versions and prioritize upgrades to Confluence 10.0+ where the feature is fully removed.
- Restrict internet exposure of REDCap, LiteLLM, and Synology DSM instances. None of these should be directly internet-accessible without VPN or zero-trust network access controls.
- Deploy WAF rules or reverse proxy ACLs blocking the
/guardrails/test_custom_codeURI on LiteLLM deployments and the XML-RPC/SOAP endpoints on Confluence. - Hunt for indicators of compromise — look for outbound connections to Cloudflare Tunnel (
*.trycloudflare.com) and Pinggy (*.pinggy.io) domains from Confluence hosts, unexpected MySQL authentication events using application credentials, and any scheduled tasks or cron entries created by the Confluence service account. - Audit LLM gateway configurations — ensure all LiteLLM endpoints require authentication, and review cloud provider billing for anomalous API usage that could indicate LLMjacking.
Strategic Guidance
Organizations should treat this KEV batch as a signal to inventory their AI/ML infrastructure with the same rigor applied to traditional enterprise software. LiteLLM and similar LLM orchestration tools are being deployed rapidly, often outside normal security review processes. Security teams need visibility into what AI tooling is running, where it's exposed, and whether it's patched. The attack surface is expanding faster than most organizations are tracking it.