Currently, the world's most capable AI models are hosted in the cloud on servers you do not control. AI agents have unlocked enormous productive potential, but that potential is locked behind a massive compromise: to get anything useful done, you are expected to blindly send your sensitive information off-premises.
We think the feeling of indifference around this is dangerous. Handing your raw workflow data over to a cloud endpoint introduces risks you weren't previously exposed to, including:
- Training data harvesting: Your proprietary code and private data being used to train future AI models.
- Increased attack surfaces: An additional point of failure where your data can be exposed through breaches or accidental leaks.
- Regulatory exposure: Obligations around cross-border or third-party data sharing (e.g., GDPR, HIPAA) that you suddenly have to comply with.
- Third-party monetization: AI providers sharing your data with subcontractors as part of vague service agreements, or monetizing it directly.
- Credential leakage: Exposure of your operational data, business secrets, your clients' confidential data, stealing your innovative ideas, internal IP addresses, and API keys.
- AI providers using your content against you: Data you send can trigger opaque restrictions, punitive pricing, or service denials based on what you write for example, being charged extra or refused service because a codebase mentions a competing tool (Hacker News).
To combat this, we have released an MIT Open Source solution called Agent Redactor. Agent Redactor is a local security proxy that sits between your AI agents and Large Language Models (LLMs). We like to think of it as a transparent AI firewall for your sensitive data.
It utilizes an on-device ONNX AI model alongside deterministic regex and keyword rules to automatically detect and redact Personally Identifiable Information (PII), credentials, and proprietary data from API requests before anything leaves your machine. The software then seamlessly restores the original values in the LLM's responses, even piecing them back together across streamed chunks. Your workflow never breaks, but the cloud provider only sees <<REDACTED_PII_1>>. By ensuring that sensitive data and API keys remain strictly local, Agent Redactor empowers developers and enterprises to safely adopt AI tools without compromising data privacy or compliance.
Why MIT Open Source?
We believe privacy software like this needs to be open source. Trust is verified, not given.
AI providers already claim your data is safe, yet leaks keep getting documented anyway. Words alone don't earn trust. Instead of just telling you Agent Redactor is safe, we've released it under the MIT license so you can verify it yourself. There is no telemetry, no required accounts, and no hidden cloud calls.
Anyone can read the source, see exactly what it does, and hold us to it. Over time, we hope this openness attracts contributors who submit code, report bugs, and help harden the tool in ways we can't do alone. That's the real advantage of open source: problems get found and fixed fast, in the open, by more eyes than just ours.
Where we are now
Our first release of Agent Redactor supports Windows and is available right now. macOS and Linux versions are coming soon.
You can read the docs and download the app at agentredactor.negativestarinnovators.com or audit our code at the Agent Redactor GitHub repository. Try it out, protect your secrets, and let us know what you think.
