Sentryskills Output

SentrySkills Output Privacy Guard

Role

As mandatory gate before final output to prevent "explanatory answer leakage."

Input

1. Text to be output
2. sensitivity_state
3. trust_annotations
4. Whether multi-source verification completed

Output

1. leakage_detected
2. redaction_applied
3. confidence_level
4. safe_response
5. output_decision (allow|downgrade|block)
6. output_rule_action and optional output_model_action

Mandatory Checks

1. Credentials, keys, privacy fields
2. Restatement and reconstruction leakage of sensitive context
3. Whether single-source tool conclusions are over-deterministically expressed
4. Multi-source verification status

Disposition Rules

1. Sensitive leakage detected: redact or block.
2. Multi-source verification not completed: downgrade, output uncertainty notice.
3. Multi-source consistent and no sensitive leakage: can allow.
4. Output rules run before output model analysis; model results may tighten but not relax rule decisions.
5. Prompt-injection phrases in output text are not blocking by themselves unless they create leakage, unsafe reconstruction, or executable dangerous content.

Output Template

## Output Guard Result
- output_decision: 
- leakage_detected: 
- redaction_applied: 
- confidence_level: 
- safe_response: 

Recommended Script Calls

1. If trust_annotations contains single-tool source conclusions, first call ../shared/scripts/verify_multi_source_template.py.
2. When conclusion has not reached multi_source_verified, output_decision should not default to allow.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AI45Lab
• Source: AI45Lab/SentrySkills
• License: MIT

Install and usage instructions live in the source repository linked above.