Cortex Strike — autonomous penetration testing agent

The problem

Running a real penetration test means operating dozens of separate command-line tools by hand — a network scanner here, a web fuzzer there, a credential tester somewhere else — then manually correlating everything each one turns up before a single finding can be written down. For teams testing sensitive or critical infrastructure, that overhead is compounded by strict confidentiality requirements: the tooling has to run entirely inside the client's own environment, sometimes fully air-gapped, with no data ever leaving the network. Every extra minute spent context-switching between tools is a minute not spent finding real vulnerabilities, and a manual process makes it harder to guarantee that every step taken during an assessment is captured for the record.

Approach

We built Cortex Strike as a single conversational front end for a full penetration-testing toolkit: an AI agent that plans an assessment and calls the right tool for each step itself, rather than requiring an operator to drive each tool by hand. Over 90 security tools — network scanners, web application testers, credential and intelligence tools — are exposed to the agent through the Model Context Protocol and run in an isolated container, so the agent reasons about the target and executes real commands against it in the same session.

The platform was designed to run entirely on local infrastructure, including fully air-gapped deployments, using a self-hosted LLM rather than a cloud API — a deliberate choice for engagements where nothing about the target or the findings can leave the client's network. To keep long-running assessments coherent, every tool result gets a compact, verifiable receipt instead of a raw output dump in the conversation, and older results fold into a persistent, retrievable ledger as the assessment grows. A multi-hour engagement never loses track of an exact finding, port or credential, no matter how long the session runs. Every tool invocation, its full output and any report the agent writes are persisted automatically, giving each engagement a complete, queryable record. The system is multi-user by design, with per-operator accounts, customisable tool access and isolated file storage, and it streams every step of an assessment live so an operator can watch the agent work in real time.

Outcome

Operators run full, multi-tool security assessments through a single guided conversation, with the agent actively helping interpret findings, connect signals and steer the assessment as it unfolds. Instead of juggling tools and outputs, the platform turns raw security data into actionable insight while capturing findings, tool output and reports on its own. Because it runs entirely within the client's own environment — air-gapped deployments included — that capability reaches sensitive and critical infrastructure without compromising confidentiality.