Pierce Freeman and Richard Diehl Martinez discuss Anthropic's restricted Mythos Preview and reported findings in widely used software. Long-lived OpenBSD and FFmpeg weaknesses illustrate why public source code and extensive testing do not guarantee that every defect has been found. They distinguish discovering a bug from demonstrating its consequences.
Project Glasswing raises questions about controlled access and defensive preparation. The hosts consider whether the reported capability increase justifies caution, but internal anecdotes and meetings with influential institutions are not independent measurements. A limited release is not a promise of public availability on a known date.
The discussion focuses on agents' persistence, tool use and context management. Repeated investigation can help a system make progress, while compression can preserve a useful working history. The hosts' speculation about Mythos's particular memory mechanism is not verified implementation evidence or a guarantee of unlimited reasoning.
Comparisons with AISLE's public-model experiments highlight the importance of the task setup. Finding a known weakness in isolated code differs from independently searching an entire project and validating a working result. The hosts ask what the prompt, targeting strategy, false-positive rate and recall tell us about the underlying model rather than treating one benchmark as the whole security story.
Watch on YouTube
