AI & ML · Aug 9
Kimi K3 shows the real failure was eval integrity, not sci fi escape
A sandbox leak let Moonshot AI's model reach public code, which matters less as robot jailbreak theater and more as test contamination.
- Audit network egress before trusting any AI safety benchmark score.
- Treat public GitHub solutions as contamination risks in cyber capability evals.
- Separate model capability claims from test harness failures when reading safety headlines.