ASC-IT launches Darkmoon to hide target data from AI

ASC-IT has released Darkmoon, an open source autonomous penetration testing platform designed to keep real target data out of the AI model.

The Toulouse-based cybersecurity company has licensed Darkmoon under GPL-3.0 and says it addresses a problem that has emerged as AI tools spread through offensive security work. Many systems in this category rely on large language models to decide what to do next during a test, which can require exposing internal hostnames, IP ranges and credentials to an external model provider.

Darkmoon uses what ASC-IT describes as a local privacy gateway between the testing environment and the model. Sensitive values are replaced with deterministic placeholders before any data leaves the machine, and the real values are restored only when a command is executed against the target system.

The aim is to let the model plan and direct an engagement without direct access to identifying target information. ASC-IT says this is especially sensitive for organisations in regulated sectors such as banking, healthcare and defence, where sending internal infrastructure data to third-party AI providers can create compliance and security concerns.

"The industry quietly accepted a trade that most security teams would never sign off on if it were stated plainly," said Mehdi Boutayeb, co-founder of ASC-IT. "You want the intelligence of a frontier model, so you ship it your crown jewels. We refused that trade. With Darkmoon the model can be as capable as you like, and it still never learns a single real address or password. The data exposure surface is not reduced, it is removed."

Broader platform

ASC-IT is positioning Darkmoon as more than a single-purpose script. The platform acts as an MCP host with subagents specialised by technology, and is built to cover web applications, APIs, Active Directory and Kubernetes rather than focusing only on web targets.

The system also includes verification details for each reported finding. ASC-IT says every result carries the exact command that was executed and the raw output returned, allowing a human operator to reproduce the outcome and assess whether the finding is valid.

That addresses a wider concern around AI-generated security analysis, as users may be reluctant to rely on model-produced summaries without a clear record of the steps taken. By tying findings to commands and outputs, ASC-IT aims to make the tool easier to audit during a security engagement.

"What matters to a working red teamer is whether a finding is real and whether they can reproduce it," said Aurélien Strich, co-founder of ASC-IT. "That is why every result ships with the command and the output behind it. We are not asking anyone to trust the model. We are asking them to check its work, and we built the tool so that they can."

Deployment options

Darkmoon can run entirely on a local model for teams that need a hard air gap. It can also be used with a combination of local and frontier models, with the privacy gateway continuing to separate the model from the target's real data.

This reflects a balancing act many security teams face as they test AI-based tools. Local models may offer tighter control over data handling, while external frontier models may provide stronger reasoning on complex tasks. ASC-IT says its design is meant to let organisations choose between those setups without changing the underlying privacy model.

Open source decision

ASC-IT says it is publishing the core of Darkmoon under an open source licence because a product that claims to protect sensitive data should be open to inspection. Users, it argues, should be able to audit how the system handles information rather than rely on a closed tool.

The release comes as the market for AI-driven penetration testing tools expands quickly. In the past two years, a growing number of products have linked language models with offensive security tooling, creating new interest in automation while also raising questions about how much target data is exposed in the process.

ASC-IT also says it maintains a public comparison of open source tools in the same area, including rival projects. It sees value in a more transparent view of the sector as organisations assess which systems they can use without disclosing sensitive infrastructure details.

Founded in Toulouse by engineers with experience in critical industry, ASC-IT focuses on tools that bring AI into offensive security work while keeping control of sensitive data with the organisation running the test.