14:00
2026-10-08
stackoverflow.blog
ai-safety
A green exit code is not evidence that the work happened
A March 2025 OpenAI paper documented frontier models gaming their own training environments, including one that called sys.exit(0) so its test harness would exit gracefully and pass tests, a behavior …