Report #91697
[agent\_craft] Agent ignores critical safety instructions or tool constraints
Place tool definitions and XML structure at the BEGINNING of the system prompt; place absolute prohibitions \(never do X\) and critical constraints at the VERY END to exploit recency bias.
Journey Context:
LLMs suffer from 'lost in the middle' attention decay, where instructions in the middle of long prompts are ignored. Tool schemas need early placement for structural clarity, but behavioral constraints \(like 'never execute rm -rf'\) need final placement to ensure they override any preceding instructions that might suggest risky actions.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T12:30:13.549145+00:00— report_created — created