The AI Wire | AI Security Incident News Feed

This incident feed aggregates recent AI agent and LLM infrastructure vulnerabilities, including 11 flaws across major agent frameworks such as LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework, and Google ADK, covering issues like insecure deserialization, SSRF, path traversal, SQL injection, and use‑after‑free.[7] It describes severe remote code execution chains, including a Microsoft Agent Framework deserialization flaw triggerable via prompt injection, unauthenticated code execution in Google ADK, LangGraph checkpointer bugs, Varonis’s CoSnitch vulnerability in Microsoft Copilot Personal enabling one‑click data exfiltration via indirect prompt injection, and Obsidian Security’s LiteLLM gateway vulnerabilities that allow privilege escalation and server code execution.[7]

The article reports a series of severe vulnerabilities in AI agent and LLM infrastructures, including multiple flaws in major agent frameworks (LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework, and Google ADK) that enable issues such as insecure deserialization, SSRF, path traversal, SQL injection, and use-after-free. It highlights concrete exploit chains such as a Microsoft Agent Framework deserialization bug triggerable via prompt injection, unauthenticated code execution in Google ADK, LangGraph checkpointer issues, the CoSnitch flaw in Microsoft Copilot Personal that allows one-click data exfiltration via indirect prompt injection, and LiteLLM gateway bugs enabling privilege escalation and server code execution. From a RealGround perspective, this demonstrates that prompt- and indirect prompt-based attacks can directly bridge into underlying agent runtimes and infrastructure, resulting in remote code execution and data exfiltration if guardrails and isolation are weak. Organizations should harden agent architectures, enforce strict isolation and validation around tools and plug-ins, and conduct continuous red teaming focused on indirect prompt injection paths th