Eliciting Harmful Capabilities by Fine-Tuning on Safeguarded Outputs

This is a ticketed event. Please register at this link.

In this talk, Talha Paracha will present insights from his latest research on using language models for improving software security ("Hallucinating Certificates", to appear at ICSE 2026).

Certificate validation is a crucial step in Transport Layer Security (TLS), the de facto standard network security protocol. Prior research has shown that differentially testing TLS implementations with synthetic certificates can reveal critical security issues, such as accidentally accepting untrusted certificates.

Paracha et al. introduce a new approach, MLCerts, to generate synthetic certificates that leverages generative language models to more extensively test software implementations. Recently, these models have become (in)famous for their applications in generating content, writing code, and conversing with users, as well as for "hallucinating" syntactically correct yet semantically nonsensical output. The authors leverage two novel insights in their work: (a) TLS certificates can be expressed in natural-like language, namely in the X.509 standard that aids human readability, and (b) differential testing can benefit from hallucinated malformed test cases. MLCerts finds significantly more distinct discrepancies between the five TLS implementations OpenSSL, LibreSSL, GnuTLS, MbedTLS, and MatrixSSL than the state-of-the-art benchmark Transcert.

​​​Event Schedule
6:00 to 6:30 - Food and introductions
6:30 to 7:30 - Presentation and Q&A
7:30 to 9:00 - Open Discussions

​​​​​​If you can't make it in person, feel free to join the live stream starting at 6:30 pm, via this link.