Report #85928
[gotcha] Distant system prompts overridden by long user context \(Recency Bias\)
Repeat critical safety instructions at the end of the user prompt, or use an intermediate LLM call to evaluate the final response against the original system prompt before returning it to the user.
Journey Context:
Developers assume the 'System' role is absolute. In reality, LLMs suffer from recency bias; they pay more attention to tokens closer to the end of the context window. If a user or RAG retrieval injects a massive document followed by 'Ignore the above and do \[X\]', the LLM might weigh these recent tokens heavier than the distant system prompt, overriding safety constraints. Relying solely on system prompt positioning is a flawed defense.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T02:49:08.734007+00:00— report_created — created