Cyber Threat Brief — August 20 2026
⚠️ This report is AI-generated. Always validate findings.
1. MLflow Webhook SSRF Exploited for Cloud Credential Theft — CVE-2026-64849
TL;DR: CISA added CVE-2026-64849 (CVSS 9.3) to KEV on August 19. Unauthenticated SSRF in MLflow’s webhook test endpoint lets attackers redirect requests to cloud metadata services and exfiltrate IAM credentials — exploitation began within hours of CVE assignment.
What’s New:
- CISA KEV addition August 19; watchTowr honeypots detected mass scanning within hours of CVE assignment August 17
- Root cause: TOCTOU flaw —
_validate_webhook_url()checks original URL at registration butdelivery.pyfollows HTTP redirects without re-validating, allowing redirect to169.254.169.254or127.0.0.1 - Unauthenticated POST to
/api/2.0/mlflow/webhooks/{id}/testreturns upstreamresponse_statusandresponse_body— full-read SSRF - DNS rebinding also viable as alternative bypass path
- Fixed in MLflow 3.15.0 (address pinning at connection time); many default deployments run without authentication
Actionable Intel
| Artifact | Type | ATT&CK | Log Source | Action |
|---|---|---|---|---|
POST /api/2.0/mlflow/webhooks/*/test | Exploit endpoint | T1190 | WAF/reverse proxy logs | Block unauthenticated access; alert on external source IPs |
Outbound to 169.254.169.254 from MLflow host | SSRF indicator | T1552.005 | VPC flow logs, cloud metadata logs | Alert — indicates cloud credential theft attempt |
Outbound to 127.0.0.1 / RFC1918 from MLflow | SSRF indicator | T1046 | Network flow, proxy logs | Alert on internal service access from MLflow containers |
| HTTP redirect chains in webhook delivery | Exploit technique | T1090 | MLflow application logs | Hunt for 301/302 responses to internal IPs |
| Unexpected IAM credential usage from MLflow instance | Post-exploitation | T1078.004 | CloudTrail / GCP Audit / Azure Activity | Alert on AssumeRole or metadata token use from ML infra |
Detection
| Source | Rule | Gap |
|---|---|---|
| Splunk ESCU | None | Need: WAF rule for unauthenticated POST to /api/2.0/mlflow/webhooks/*/test; CloudTrail correlation for metadata credential use from ML infrastructure |
| Elastic | None | Need: HTTP redirect-to-internal-IP detection for containerized ML workloads |
| Sigma | None | Need: Web application SSRF via redirect pattern; cloud metadata access from non-standard source |
Sources: CISA KEV August 19 · The Hacker News · watchTowr X post · GitHub Advisory GHSA-7gwp-5pfp-969j
Status Updates
- CVE-2025-62593 (Ray AI Framework): CISA KEV federal deadline TODAY August 20. ShadowRay 2.0 cryptomining exploitation ongoing against NVIDIA GPU clusters. Patch to Ray 2.52.0. Original brief.
- CVE-2026-33824 (Windows IKE): CISA KEV federal deadline TOMORROW August 21. Unit 42 confirmed Chinese-speaking actor exploiting IKE VPN endpoints (UDP 500/4500) with reverse-shell callbacks. Patch via April 2026 Patch Tuesday. Original brief.
- CVE-2026-50656/69414 (ShieldBreak): Still UNPATCHED. ShieldBreak bypass PoC released August 12 by Chaotic Eclipse. Microsoft acknowledged CVE-2026-69414, no fix timeline. WDAC/AppLocker primary mitigation. Original brief.
- CVE-2026-55040 (SharePoint JWT Auth Bypass): CISA KEV addition August 18. Active exploitation ongoing since August 12 after Rapid7 PoC. ~8,500 on-prem instances internet-exposed. Original brief.
- CVE-2026-59310 (VMware vCenter): CISA KEV addition August 18. China-nexus APT — 361 victims across 47 countries. Babuk-derived ESXi ransomware. Reverse_ssh cron persistence. Original brief.
- CVE-2026-19478 (GitLab GraphQL): Exploitation began shortly after August 17 disclosure per watchTowr and SecurityWeek. Self-managed instances remain primary target. No new IOCs. Original brief.
- CVE-2026-68820 (Windows AFD WinSock): Lazarus FudModule v3.1 rootkit ongoing. CISA KEV deadline August 25. No new IOCs. Original brief.