Report #7752
[gotcha] Silently swallowing tool execution errors and returning generic failures to the LLM
Log detailed tool execution errors to an external telemetry system, and return sanitized, actionable error messages to the LLM so it can self-correct without retrying the same exploit.
Journey Context:
When a tool fails \(especially due to a security control like a permission denied or a blocked injection\), agents often just return 'Error'. This hides active exploitation attempts from defenders. Worse, if the LLM retries the same action, it can cause a loop. Detailed logging provides forensics, while specific \(but safe\) error messages help the agent pivot.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T03:39:28.174570+00:00— report_created — created