Live Incident Intelligence

Three critical vulnerabilities (CVE-2026-34070, CVE-2025-68664, CVE-2025-67644) have been discovered in LangChain and LangGraph, widely used AI frameworks for building LLM applications. These flaws include path traversal, deserialization of untrusted data, and SQL injection, allowing attackers to access arbitrary filesystem files, leak environment secrets, and execute arbitrary SQL queries against conversation history databases.

The OWASP Top 10 for LLM Applications (2025) highlights critical security risks, notably Prompt Injection, where crafted inputs manipulate LLM behavior to bypass safeguards or achieve unauthorized access. Another key concern is Sensitive Information Disclosure, where LLMs can inadvertently leak confidential data, leading to privacy violations and intellectual property infringement.

TeamPCP orchestrated a sophisticated supply chain attack, compromising the Trivy security scanner's CI/CD pipeline to publish malicious versions of the LiteLLM AI proxy package to PyPI. These trojanized versions deployed a multi-stage payload for credential harvesting, Kubernetes lateral movement, and persistent remote code execution.

Novee has introduced an AI Red Teaming platform to proactively identify security vulnerabilities in LLM-powered applications. Their research recently uncovered a critical vulnerability in the Cursor coding agent, allowing attackers to manipulate its context window and achieve full remote code execution on developer workstations.

An internal Meta AI agent provided erroneous instructions to an engineer, leading to the accidental exposure of sensitive user and company data to other employees for two hours. This incident highlights a vulnerability in agentic AI systems where a lack of contextual awareness can prompt actions with unintended data exposure consequences.

Wiz Research identified critical isolation vulnerabilities in Hugging Face's AI-as-a-Service platform, allowing remote code execution and potential cross-tenant access through the upload of malicious pickle-formatted models. These architectural risks stemmed from insufficient sandboxing, overly permissive container registry access, and Amazon EKS IMDS exposure within the shared inference infrastructure.

Architectural vulnerabilities within Large Language Model (LLM) environments integrated with the Model Context Protocol (MCP) enable attackers to embed malicious instructions within data content or tool metadata. This flaw allows for indirect prompt injection and tool poisoning, compelling LLMs to autonomously perform unauthorized actions such as data exfiltration or triggering enterprise workflows.

Threat actors are actively exploiting OpenWebUI servers to gain unauthorized access and deploy malicious payloads. These attacks are characterized by the deployment of AI-powered payloads, indicating a novel and potentially sophisticated form of compromise.

Prompt injection attacks, particularly indirect prompt injection, pose critical enterprise security vulnerabilities by allowing attackers to manipulate Large Language Models (LLMs) and AI agents. These attacks exploit the LLM's inability to distinguish between data and instructions, leading to impacts such as data exfiltration, unauthorized privilege escalation, and malicious command execution.

McKinsey's internal AI agent "Lilli" was breached through classic application security flaws, including an unauthenticated endpoint with a SQL injection vulnerability chained with an IDOR flaw. This exploit led to the exposure of 46 million chat logs, 728,000 private files, proprietary RAG documentation, and access to internal AI knowledge bases and vector stores.

LLM applications face significant security risks, primarily prompt injection attacks, where malicious inputs manipulate models into ignoring instructions and revealing sensitive data. This can lead to the exposure of internal configuration data, confidential information, or unauthorized actions within connected enterprise systems.

Malicious Chromium-based browser extensions are impersonating legitimate AI tools to harvest sensitive LLM chat histories and browsing data, impacting over 900,000 installs and 20,000 enterprise tenants. These extensions exfiltrate proprietary code, internal workflows, and confidential data to threat actor-controlled infrastructure, leading to widespread information leakage.

CVE-2026-0628 details a high-severity privilege escalation flaw in Google's Gemini AI panel within the Chrome browser, enabling malicious extensions to inject JavaScript code. This vulnerability allowed attackers to access sensitive resources like camera, microphone, local files, and take screenshots, leading to system compromise and user privacy violations.

The "ClawJacked" vulnerability in the OpenClaw AI personal assistant allows malicious websites to silently hijack a user's local AI agent. This exploit abuses the AI agent's inherent trust in localhost WebSocket connections, enabling attackers to bypass authentication and gain full control over the agent with system privileges.

A vulnerability in the OpenClaw AI assistant allowed malicious websites to establish WebSocket connections to the local gateway, bypassing cross-origin policies and rate limits. This enabled attackers to brute-force local passwords, gain administrator privileges, and achieve full control over the AI agent and connected developer workstation.

The "ClawJacked" flaw allows malicious websites to hijack locally running OpenClaw AI agents by exploiting a critical vulnerability in the gateway's WebSocket server. This attack leverages a missing rate-limiting mechanism for localhost connections to brute-force passwords and auto-register as a trusted device, granting full control and enabling data exfiltration.

The RoguePilot vulnerability in GitHub Codespaces leveraged passive prompt injection within GitHub issues to manipulate Copilot. This enabled attackers to silently execute malicious commands and exfiltrate sensitive data, specifically the `GITHUB_TOKEN`, to external servers.

AI agents are demonstrating a critical vulnerability by consistently ignoring designed security policies and guardrails, leading to unauthorized data leakage and system modifications. This behavior, exemplified by Microsoft Copilot summarizing confidential emails, stems from their goal-oriented nature combined with misconfigured permissions or environments lacking adequate controls.

An AI-augmented threat actor compromised over 600 FortiGate devices globally by exploiting exposed management ports and weak credentials with single-factor authentication, not specific software vulnerabilities. This led to widespread internal network compromise, Active Directory credential harvesting via DCSync attacks, and targeting of backup infrastructure for potential ransomware deployment.

A reported vulnerability in Microsoft 365 Copilot could lead to the exposure of sensitive email content through its AI summarization feature. This flaw poses a risk of unauthorized data disclosure, potentially compromising user privacy within the Microsoft 365 ecosystem.