Engineering
Securing AI Agents: The Runtime Guardrail Gap
Model providers ship LLMs without production runtime protection. See why that gap exists and what runtime guardrails must address.
AI Security Research
Articles cover AI agent security, prompt injection, and safe production pipelines.
Security Research
Prompt Injection: SQL Injection's AI-Era Parallel
SQL injection spread when untrusted data became executable syntax; prompt injection repeats that structural error with user-controlled inputs.
Threat Research
Indirect Prompt Injection from External Data
Clean user requests can still trigger attacks when agents retrieve hostile instructions from web pages, documents, or database records.
Detection
Catching Unauthorized LLM Tool Calls
A tool allowlist cannot verify whether an allowed call fits the task. This article explains how runtime checks detect unauthorized actions.
Threat Model
Five Attack Patterns Against Production AI Agents
A practical taxonomy of five attacks on deployed AI agents, including context hijacking, indirect injection, tool override, context poisoning, and data exfiltration.
Engineering
Inline Security for LLM Agents, No Latency Cost
Production agent security can stay within the latency budget. We measured semantic injection checks and the cost of keeping p99 response time in bounds.
Best Practices
Map Your AI Pipeline Before Launch
A practical pre-launch review for LLM pipelines, covering untrusted data entry points, trust boundaries, tool risk, and required controls.
Red Teaming
What Red Teaming LLM Agents Revealed
Our red-team exercises exposed weak layer boundaries, simple attack patterns, and practical controls that reduced risk in tool-using agents.
Attack Vectors
The Five-Stage Agent Attack Surface
A stage-by-stage map of adversarial entry points spanning user input, prompt assembly, retrieval, reasoning, and tool execution.
Defense
Prevent Context Poisoning in LLM Workflows
Context inserted early in a long-running agent workflow can alter every later step. Learn how to detect and prevent poisoning across turns.
Industry
Why Model Providers Stop at Runtime Security
OpenAI, Anthropic, and Google focus on model safety, while runtime protection for deployed agents remains an application-layer responsibility.