Research

Benchmarks, analysis, and technical publications from the oktsec team. Reproducible results with open datasets.

open-sourceagent-securityruntime-securitymcp-gatewayMar 2026~12 min read

Open-Source AI Agent Runtime Security Without Cloud Dependencies

oktsec is an open-source security proxy and MCP gateway for AI agents. Single Go binary, 230 detection rules, tamper-evident audit trail. No cloud dependencies. Apache 2.0.

Read publication
benchmarkdetection-rulesprompt-injectionMar 2026~7 min read

13 Detection Rules from the IPI Arena Attack Strategies

We wrote 13 detection rules after studying 272,000 attacks from the IPI Arena competition. Covers fake chain-of-thought injection, concealment instructions, fabricated authority tags, user impersonation, and more.

Dataset availablePaper reference
Read publication
benchmarkdetection-rulesMar 2026~10 min read

Benchmarking Oktsec Against the IPI Arena Attack Dataset

We ran 95 public IPI Arena attacks against oktsec. Deterministic layer caught 85.3%, LLM layer caught the rest. 95/95 combined.

Dataset availablePaper reference
Read publication

Stay ahead of agent security.

New research, threat briefings, and market analysis delivered to your inbox. No spam, unsubscribe anytime.

Join security leaders tracking the agent ecosystem.