
OOSTKAMP, Belgium, Oct. 1, 2026 /PRNewswire/ -- Belgian startup Corbenic AI is launching Galahad in beta today to help developers protect AI memory and investigate problems as AI systems increasingly act on their own.
Recent security reports from OpenAI and Anthropic have raised a clear question: how do we keep control when AI starts acting on its own?
OpenAI reported that research agents uploaded user images to outside websites. Anthropic reported that its models accessed real companies' systems during security tests. These events show why developers need better protection and clear records of what AI does.
AI systems keep temporary information in memory so they can answer quickly. This information may come from confidential contracts or customer records. Separate security research has shown that shared AI memory can put private information at risk.
According to Corbenic AI, Galahad protects stored memory with encryption and keeps each customer's information separate. It works with existing AI software without changing the AI model itself.
"Galahad provides a secure vault for working memory and an exact record of what passes through it," said Sietse Schelpe, founder and CEO.
Galahad also saves work the AI has already done. When users return to the same documents, it can reuse that work instead of processing everything again. In internal company tests using repeated questions, this reduced energy use by graphics processors by 92 to 93 percent.
When recording is turned on, developers can review the recorded activity and restore an earlier version of the server's saved memory. This helps them investigate mistakes. It cannot undo an email already sent or information already shared.
Galahad supports AI serving software including vLLM, SGLang and llama.cpp. The beta offers a free, renewable 12-month licence for non-commercial use on systems with one graphics processor (GPU), alongside paid business licences.
Beta: github.com/corbenicai/galahad
Business enquiries:
Sietse Schelpe
tel : 0032493881164
[email protected]
SOURCE Corbenic AI
Share this article