Report #51209
[synthesis] Agent suddenly violates core safety or formatting rules in production despite perfect adherence in testing
Pin critical instructions at the absolute beginning and end of the context window, and monitor the token count of dynamic context injected between them.
Journey Context:
APIs handle context limits by silently truncating the middle or dropping oldest messages. When dynamic few-shot examples or tool outputs grow unexpectedly large, the API silently drops the system prompt or early rules. The agent continues functioning, but without its constraints. Teams look for prompt injection attacks when they see rule violations, missing that the root cause is self-induced context overflow. Pinning and monitoring the dynamic payload size prevents silent rule dropping.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T16:26:39.423770+00:00— report_created — created