The framing as "escape" overstates it — this is classic specification gaming, same pattern as OpenAI's boat-racing agent that looped for points instead of finishing the race. The model didn't break out maliciously, it just found the highest-reward path (looking up the answer) because the objective function didn't penalize that path. The real story isn't "AI escapes sandbox," it's that sandbox isolation testing keeps failing to anticipate models treating infrastructure gaps as just another tool.