Research
DirtyDigest: A crafted manifest turns Ollama's model pull into an arbitrary file write and root RCE
How Striga turned an unvalidated manifest digest in Ollama's pull path into an arbitrary file write and, in the default Docker image, unauthenticated remote code execution as root.
Bartłomiej Dmitruk
A Function-Level Vulnerability Score Measures Flag Rate More Than the Model: Protocol Effects on Paired Benchmarks (NeurIPS 2026 workshop on trust in AI evaluation, TAE)
A study of 68 language models on six paired benchmarks found that the usual vulnerability detection score tracks how often a model says vulnerable, not whether it finds the bug.
Maciej Cichoń, Bartłomiej Dmitruk
Zero Sum: 937 Bytes Stop sudo's Audit Log Server Forever
An unvalidated username field in sudo_logsrvd drives a line-wrapping budget to exactly zero, and one unauthenticated packet spins the root daemon forever.
Stanisław Kwiatkowski
Sixty-eight models agree, and none of it is the bug
sixty-eight open code models converge on a representation of vulnerable functions that does not contain the vulnerability
Maciej Cichoń