A compaction summary is a condensed version of earlier conversation or task state used when an AI session grows too long. Instead of keeping every token, the system preserves the most important details so the model can continue working with less context. This is useful for long-running agents, but it also turns the summary into a security boundary.
In cyber security, compaction summaries matter because any untrusted instruction or sensitive data that survives compression can affect later actions. Attackers may try to inject malicious guidance into text that gets summarized, causing the model to keep following hidden instructions after the original source is gone. Defenders reduce this risk by limiting what is preserved, separating user content from trusted state, reviewing summaries for unwanted directives, and logging high-risk compaction events. Treat the summary as privileged state, not just a convenience.