agents

1 Article

What survives compaction: how an LLM agent loses its rules and launders other people's instructions

An analysis of the risks of context compaction in LLM agents (Claude Code, Codex, the Anthropic and OpenAI APIs): the trust boundary that disappears in the summary, the rules that get lost, the stale copy that gets re-injected and the encrypted item nobody can audit. With a reproducible proof of concept on five local models and Claude Code.