Agent red-teaming
We probe AI agents and LLM applications the way a real adversary would — prompt injection, tool abuse, privilege escalation across agent boundaries, and data exfiltration through the surfaces nobody thought to lock down.
Agentic Security
AgentPentest
AI-driven red-teaming for AI systems and the networks they run on — autonomous agents that find the flaws before an attacker does.
What we do
Offense, automated.
Malicious traffic analysis
Our tooling ingests live and captured network traffic, isolates hostile behaviour from noise, and reconstructs what an attacker actually did — turning raw packets into an attack narrative you can act on.
Detection engineering
We generate and validate detection rules for emerging threats, then test them against synthesized attack traffic — so what ships to your sensors is proven to fire on the real thing and stay quiet on everything else.
About
Built for a threat surface that didn't exist last year.
AgentPentest is a security research and tooling company working at the point where AI systems become attack surface. Autonomous agents now read untrusted input, call tools, and act on live infrastructure — and the testing methods built for conventional software don't reach them.
We build agentic offensive tooling: systems that plan, probe, and adapt on their own, running continuously instead of once a quarter. The same engine that breaks a target feeds detection engineering on the other side — every finding becomes a validated rule. Offense and defense, closing the loop.