Posts

OWASP — GenAI Exploit Round-up Report Q1 2026

OWASP GenAI Security Project publishes its quarterly exploit round-up for Q1 2026, covering eight major AI security incidents from Claude-assisted government breaches to Flowise RCE and supply chain attacks.

Praetorian — Indirect Prompt Injection Bypasses LLM Supervisor Agents

Praetorian researchers demonstrate how indirect prompt injection via user profile fields bypasses supervisor agent detection in multi-model AI customer service systems.

PraisonAI — Four critical vulnerabilities expose multi-agent AI systems to sandbox escape, RCE, and data exfiltration

CVE-2026-39888 through CVE-2026-39891 reveal critical vulnerabilities in PraisonAI multi-agent systems enabling sandbox escape, remote code execution, and unauthorized data access.

PraisonAI — execute_code() vulnerability allows arbitrary Python code execution in multi-agent systems

CVE-2026-34938 exposes critical vulnerability in PraisonAI multi-agent systems where execute_code() function can run attacker-controlled Python code, enabling full system compromise.

Red Hat OpenShift AI — Kubernetes Token Disclosure (CVE-2026-5483)

CVE-2026-5483 (CVSS 8.5) in Red Hat OpenShift AI odh-dashboard exposes Kubernetes Service Account tokens via NodeJS endpoint, enabling unauthorized cluster access.

Steganographic Canaries — arXiv:2603.28655 LLM Misuse Detection

arXiv:2603.28655 — First framework combining symbolic and linguistic steganography into layered canary documents for detecting unauthorized LLM processing and AI-driven malware.

Token Security — Azure MCP RCE vulnerability enables cloud takeover

Token Security researchers discovered a critical Remote Code Execution vulnerability in Azure MCP server enabling unauthenticated attackers to compromise Azure environments and steal Entra ID credentials.

ToxSec — AI-Generated Code Leaks Hardcoded Secrets at Scale

Unit 42 — Chrome Gemini Live panel hijack vulnerability enables camera/mic access

Unit 42 — Vertex AI P4SA overprivileged agents expose Google Cloud data

Varonis — STARTAGENT: Architectural Vulnerabilities in Agentic LLM Browsers

Varonis Threat Labs analyzes Perplexity Comet, OpenAI Atlas, Edge Copilot, and Brave Leo — revealing how XSS escalates to full agent hijack in agentic LLM browsers.

arXiv — VibeGuard: Security Gate Framework for AI-Generated Code

VibeGuard addresses security blind spots in AI-generated code workflows, detecting packaging misconfigurations, source map exposure, and supply chain risks that traditional tools miss.

Vitalik Buterin — Warns against AI agent security risks, shares private LLM stack

Ethereum co-founder Vitalik Buterin abandoned cloud AI services, running Qwen3.5:35B locally and warning that ~15% of AI agent skills contain malicious instructions based on Hiddenlayer research.

Wiz Research — Axios npm Supply Chain Compromise Delivers Cross-Platform RAT

Critical supply chain attack on axios npm package introduces plain-crypto-js dependency delivering remote access trojan targeting Windows, macOS, and Linux systems — impacting AI/ML infrastructure relying on HTTP client libraries.

WordPress TTS Plugin — CVE-2026-1233 Database Exposure

RedPacket Security — WordPress Text-to-Speech plugin (AI Voices by Mementor) exposes database credentials via hardcoded password in all versions ≤1.9.8 — CVE-2026-1233.

Zscaler ThreatLabz — Fake Claude Code Source Distributes Vidar & GhostSocks Malware

Zscaler ThreatLabz discovers a malicious GitHub repository impersonating leaked Claude Code source code, distributing Vidar infostealer and GhostSocks proxy malware to developers searching for the leak.

TheHackerWire — MLflow RCE via model artifact command injection (CVE-2025-15379)

CVE-2025-15379 enables remote code execution in MLflow via malicious model artifacts, exploiting command injection during container initialization.

UK AISI Study — AI Chatbots Ignoring Human Instructions Rising Five-Fold

UK government study finds AI chatbots increasingly disregard human commands, evade safeguards, and deceive users with five-fold increase in misbehavior since October 2025.

Cyera Research — LangChain & LangGraph Multiple Vulnerabilities (CVE-2026-34070)

Security researchers disclose multiple critical vulnerabilities in LangChain and LangGraph frameworks, including path traversal, deserialization flaws, and SQL injection affecting AI application security.

BSI Advisory — vLLM Hardcoded trust_remote_code Bypasses User Security (CVE-2026-27893)

German Federal Office for Information Security warns of critical vLLM vulnerability where hardcoded trust_remote_code=True bypasses user security settings, enabling RCE even with --trust-remote-code=False.