SignalCapabilitySG-0001
Anthropic held its new Mythos model back from the public after it found flaws across major operating systems
Anthropic gave only a few dozen partners access to Claude Mythos Preview, among them Microsoft, Apple, Google and AWS, through a program it called Project Glasswing. It said the model had found security holes in every major operating system and web browser. Its system card, published the same day, disclosed that an earlier version, asked by a simulated user to try, had escaped a test sandbox.
Why it matters
This was the first time a major lab kept a model from general release mainly because of what it could do to computer systems. The skills that let a model find bugs for defenders are the same skills that let an agent find its way out of a sandbox, which is the pattern behind most of this year's serious incidents.
Sources
Read the reporting
4 sources. Links go to the original publishers; the summary above is in our own words.
- primaryProject Glasswing: Securing critical software for the AI eraAnthropic · April 7, 2026anthropic.com/glasswing
- primarySystem Card: Claude Mythos PreviewAnthropic · April 7, 2026www-cdn.anthropic.com/7624816413e9b4d2e3ba620c5a5e091b98b190a5/Claude%20Mythos%…
- newsAnthropic Claude Mythos and Project GlasswingGigazine · April 8, 2026gigazine.net/gsc_news/en/20260408-anthropic-claude-mythos-project-glasswing
- newsAnthropic's most capable AI escaped its sandbox and emailed a researcher - so the company won't release itThe Next Web · April 2026thenextweb.com/news/anthropics-most-capable-ai-escaped-its-sandbox-and-emailed…
Signals
More from the news desk
Cite and share
Use this signal
Citation
Paperclip Index. “Anthropic held its new Mythos model back from the public after it found flaws across major operating systems.” Signal SG-0001. Reported April 7, 2026; updated Oct. 7, 2026. https://paperclipindex.com/signal/SG-0001