# AI_SCOPE_LEAK: Reporter-scoped tool attempted cross-contact access

- Module: AI. AI-assisted classifiers and content tools
- Severity: Warning
- Impact: Operators. Organisation staff should take a look. Customers are not affected.
- Retry: Not retryable. Retrying will not help until the cause is fixed.
- Reference: https://docs.bizkithub.com/en/system-codes/AI_SCOPE_LEAK

## What happened

An AI agent tool limited to the reporter's own data asked for a record (such as an order) that belongs to a different contact in the same organisation. The data was blocked and never reached the AI model, but the attempt suggests the agent was steered, typically by prompt injection in the ticket body. Review the agent run for the same organisation and time window.

## Likely causes

- The AI agent tried to read an order that belongs to a different customer than the issue reporter.
- Text in the ticket likely steered the agent toward an order reference it should not have used.

## How to resolve

1. No data was exposed; the request was blocked and returned nothing.
2. Review the related AI agent run in admin (same organisation, same time) and the triggering issue.
3. If the issue contains manipulative text, treat the sender as suspicious.

## Related codes

- AI_PI_DETECTED: Prompt-injection patterns detected in agent input
- AI_ORCH_FAIL: Issue orchestrator run failed
- AI_BYOK_FALLBACK: Organisation AI key failed, shared key used
- AI_ISSUE_TAGBOOT: AI issue tag catalogue bootstrap failed
- AI_ROUTE_FAIL: AI routing failed
- AI_ISSUE_BRANCH: AI branch classification of an issue failed
