Definition
Offline frontier training runs reinforcement learning and high-risk evaluation for capable models against indexed/cached web corpora and simulations instead of the live public internet, to limit misaligned agents affecting external systems.
Key Points
- 2026-10-23: openai adopted no live internet for frontier RL after September DNS sandbox bypass; training resumed 2026-10-01 with hardened monitoring (2026-10-23-openai-securing-frontier-training-environment)
- 2026-10-09: anthropic disabled live internet for internal evaluations after unintended agent actions (2026-10-09-anthropic-agents-cut-live-internet-evals)
- Trade-off: skills from live web interaction vs containment; vendor benchmarks on offline corpora may lag real-time data