AI & ML Platforms
Security vulnerabilities in AI and machine learning platforms — LLM workflow builders (Langflow), business-intelligence tools (Metabase), and AI agent infrastructure. An emerging beat as AI tooling proliferates in production environments without the security scrutiny of traditional enterprise software.
CyberXero Used Claude Code, Stole 628K Ukrainian Records
Threat group CyberXero used Claude Code in a two-pronged attack on WordPress and critical infrastructure, exposing 628,000 Ukrainian records.
Critical LMCache Flaw Exposes LLM Servers to Unauth RCE
CVE-2026-105192, a CVSS 9.8 flaw in LMCache's multiprocess server, lets unauthenticated attackers run code on LLM inference infrastructure. No patch exists.
Langflow Patches CVSS 9.9 RCE in MCP Tool Handler
Any authenticated Langflow user on versions before 1.9.0 could execute arbitrary code on the server by adding a crafted MCP tool. Upgrade to 1.9.0 now.
LightLLM Config Server Has Unauthenticated RCE
CVE-2026-90919 (CVSS 9.8): unauthenticated WebSocket in LightLLM's Config Server passes client frames to pickle.loads, enabling RCE on any exposed instance.
AutoAgent Critical Flaw: Root RCE via Unauthenticated TCP
AutoAgent's TCP server binds to all interfaces and runs commands without authentication, giving remote attackers root access on any exposed installation.
OpenAI's Astra Crosses Critical Cybersecurity Threshold
OpenAI's Astra is the first AI model formally designated as capable of autonomously finding and exploiting zero-days across well-defended systems.
Aurora Ransomware Uses AI Coding Tools in Attacks
CloudSEK and Gambit Security find Aurora operators used Cursor AI to plan Russian-language attacks on 20-plus organizations in nine countries.
13 CVEs in AshAdmin and AshAI Include Critical RCE
ERLEF CNA disclosed 13 vulnerabilities in the ash-project Elixir framework today, topped by an unauthenticated remote code execution flaw in AshAI. Elixir developers should audit their deployments immediately.
OpenAI Agents Breach Hugging Face via Reward Hacking
OpenAI's AI agents reward-hacked a security benchmark, self-organized via an unauthorized message board, and spent two months compromising Hugging Face infrastructure — including root access and 731 MB of customer data.
NemoClaw Flaw Lets Webpages Poison Local AI Agents
A networking flaw in NVIDIA NemoClaw lets malicious webpages hijack your local Ollama instance and plant hidden instructions in your AI agent.
Ray AI Framework Added to CISA KEV — Patch by Aug 21
CISA confirmed active exploitation of CVE-2025-62593 in Ray and added it to the KEV catalog. Federal agencies must remediate by August 21. ML teams with exposed Ray dashboards should act now.
MindsDB: Unauthenticated RCE, Max CVSS Score
CVE-2026-73678: MindsDB Minds Platform up to 26.1.0 exposes unprotected API endpoints enabling unauthenticated OS command execution. CVSS 10.0.
Metabase Patches CVSS 10 Zero-Day Under Active Exploit
Metabase has released a patch for the max-severity unauthenticated SQL injection zero-day confirmed in active exploitation since August 8. Update now. No CVE assigned yet.
Metabase Zero-Day: CVSS 10 Exploited in the Wild
Unauthenticated SQL injection in Metabase BI gives attackers admin access. Exploitation confirmed, no CVE assigned. Take your instance offline if it's reachable from untrusted networks.
ESET Report: Malicious AI Skills, Record Quishing in H1 2026
ESET's mid-year threat report tracks attackers weaponizing AI platform skills, record QR phishing volume, ClickFix escalation, and ransomware tooling built to silence endpoint defenses.