Recently, SentinelLABS took a real investigation, one of their own the incredible findings of fast16, a 2005 Windows sabotage implant aimed at nuclear-simulation software, and rebuilt it as an eight-stage benchmark for frontier reasoning models. The test was built to test frontier models, not on whether they can untangle the puzzle, but how do they reason and adapt.  

Can a frontier model maintain a trustworthy investigation as new evidence repeatedly proves earlier conclusions wrong? Of six model families tested, two completed it, and even they made real errors along the way, including one case where an automated verification suite reported "no failures" on a conclusion that was, in fact, incorrect. 

Date: Thursday, 3 September, 2026

Time: 1:00 p.m. BST / 2:00 p.m. CEST

Gabriel Bernadett-Shapiro who led the research, joins us to walk through what actually happened inside these runs, what separated the models that recovered from the ones that stalled, and what it means for anyone deciding how far to trust AI inside their own security operations.