About us
A model does not break, it drifts: it caves under pressure, turns sycophantic, misreads the person in front of it. We read a model from its output, and from its internal layers where we have access, and turn that behaviour into numbers mapped to the research literature. Then we correct what drifted.
Read. PSA reads how a model behaves, with no access to its weights or training data needed: whether it gives in under pressure, flatters, overclaims, or follows a command hidden in a document it was asked to read.
Tune. Our tune-kit corrects what the reading found and adds the skills a model lacks, such as telling the user's request apart from an instruction planted in the data.
A reading without a correction is only a meter. We do both, and we publish what we measure.
Co-founder and CEO
Dr.-Ing. in cybersecurity, summa cum laude, with three best paper awards. 18 years of R&D in academia and industry, leading more than 10 international R&D teams and projects. Formerly at Juniper Networks and in Silicon Valley, and on the Executive Steering Group of the MHinCS Foundation.
Co-founder and CTO
About 25 years in cybersecurity, CISSP. Author of the Cybersecurity Psychology Framework, of the book Mastering Machine Learning, and of 2 arXiv and more than 10 SSRN papers on model behaviour. He built and trained PSA's readers and its correction kit. He founded and exited two companies; this is his third.
Founding team, go-to-market and compliance
15 years in compliance and ISO standards. She maps our evidence to the EU AI Act.
Advisor
Former CISO and angel investor, on the founding team of Bit9 (now Carbon Black).
Advisor
Einstein Professor at TU Berlin, Security in Telecommunications.
Silicon Psyche Labs is based in Berlin. For customers, partners and investors: book a call or write to us.