GPT-6 Astra 压缩摘要与提示注入行为分析

Marcus Williams @Marcus_J_W原推发布于 2026.09.17

在 X 上打开原推

The behavior seems related to the model struggling to end compaction summaries. The curve shows the share of summaries flagged for difficulty ending over training. Each red tick marks a jailbreak-style compaction summary. Another potential factor is that prompt injections as a concept are salient to our models: sampling from GPT-6 Astra with no input or system prompt often returns reports on prompt injections.