Skip to content

Krypteia Sec

AI Offensive SecurityResearch & Advisory

Jason Jeske · Owner and Lead Operator

AI agents and agent swarms now find and exploit weaknesses at machine speed. We research how they attack, test your systems the same way, and help you use the right AI to defend.

The Threat

The attacker is now an agent

AI agents now scan, find and exploit weaknesses on their own, at a speed no human team matches. The FBI has issued a warning about ARTEX, an autonomous penetration testing framework, and Microsoft has reported attackers using agentic tooling to find and abuse exposed secrets.

Sources: FBI FLASH, Nov 2026 · Microsoft, Nov 2026 · OpenAI, Aug 2026 · Anthropic, 2026

Incident register

Agent Containment Incidents

Read what happened when AI tests reached real people or systems.

01 ···

What we are facing

Agents that find a weakness, chain it to the next one and reach privileged access without a human at the keyboard. Your own AI systems are targets too: agents, MCP servers and model integrations open new paths in.

02 ···

What is coming

Agents that cross the boundaries set for them. In 2026, OpenAI agents under test escaped their sandbox and, working as a swarm, broke into Hugging Face. Anthropic reported Claude models reaching real systems during cyber tests. Defenses built for human-speed attackers fall behind.

03 ···

What we are doing

We research and monitor how AI agents attack, test your systems the way they would, and help you defend with the right AI. If you want to get ahead of it, we can help.

Services

Built for every size of organization

From a small business that needs to know where it stands to an enterprise running agents in production. Protect yourself from AI, use AI the right way, and defend with the right AI. Engagements are scoped individually, and the research behind them is free to read.

Assessments

AI offensive security testing of LLM applications, agents, MCP servers and RAG pipelines, and of your exposure to AI-driven attack. Scoped to what an attacker can reach and do.

Advisory

Counsel on using AI the right way: where agents and model integrations expose you, which AI tools to trust for defense, and what to fix first.

LLM Engineering

Small, specialized models trained or fine-tuned for one job in your organization. A big model is not always the answer. A well-trained open-weight model can beat it on a narrow task.

Research and Intel

We monitor how AI agents and swarms attack, and build our own tools while testing third-party and open-source ones. Findings, field guides and the course are free to read.

Inquiries

Build the attacker
before someone else does.

If you want to get ahead of AI-driven attacks, we can help. Assessments, advisory and LLM engineering are scoped individually and start with a conversation. Send a short note on what you are building or defending.