RSK-4318 · Safety
Toxicity filter false-negative on regional languages
Classifier recall drops to 0.61 on Kannada and Marathi transliterations, allowing harsh phrasing through.
medium
Exposure score
12/25
likelihood 4 × impact 3
Remediation SLA
20 days
detected 2026-07-10 11:02
Accountable owner
Customer Experience
named under the AI risk charter
Affected agent
Complaints Triage Agent
AGT-1133 · Customer Experience
Investigation
What the sweep found and how it was corroborated
Classifier recall drops to 0.61 on Kannada and Marathi transliterations, allowing harsh phrasing through.
Recall
0.61
Target
0.90
Languages
4
Exposure position
Placed at likelihood 4, impact 3 on the enterprise 5×5 matrix.
Remediation plan
Tracked to closure
- Add multilingual eval set
- Route low-confidence turns to human
Case metadata
- Risk ID
- RSK-4318
- Category
- Safety
- Detected
- 2026-07-10 11:02
- Owner
- Customer Experience
- SLA
- 20 days
- Agent record
- AGT-1133