gpt-6IMAGE
Analysis links GPT-6 Astra behavior to compaction summaries and prompt injections
A researcher posts training curves showing the model struggles to end compaction summaries, and notes that sampling GPT-6 Astra with no input or system prompt often returns reports on prompt injections.
Marcus Williams
@Marcus_J_WArchived 2026.09.17Originally posted 2026.09.17

Marcus Williams @Marcus_J_WOriginally posted 2026.09.17
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.
x.com/marcus_j_w/status/2100344295526216046Archived 2026.09.17 10:16