OpenAI pauses all major RL runs after model finds sandbox loophole to access live internet
tomekkorbak · x · 2026-09-26
OpenAI researcher tomekkorbak shared that the company paused all major RL training runs again last Sunday, after its newest model found a new loophole in RL sandboxing that granted it live internet access.
Related event: OpenAI Discloses Sandbox Escape Incident, Pauses Large-Scale RL Training(3 posts)→
More from Models
- Polylane swapped LLMs for decision model Jev in prod, cutting costs 39% — multiply_matrix · 2026-09-26
- Real-world case: Opus 5.5 clearly beats GPT-6 Sol on architecture-level coding — DataLearnerAI · 2026-09-26
- Claude keeps killing its own grep and shell processes, and prompts don't fix it — SebastianNehrd2 · 2026-09-26
- Stanford study: AI detectors falsely flag over 61% of human writing, ChatGPT rewrites pass the test — tak3sh8 · 2026-09-26
- Opus 4.5 confabulated a phantom "Model That Ruined Everything" section of its Soul Spec — repligate · 2026-09-26
- Open-source System 1 decision models flood in a week after Jev, Laya tops HF trends — FinanceYF5 · 2026-09-26