Fake bank chatbot leaked data in 54 of 55 attacks; one gateway cut it to zero
its_vayishu · x · 2026-11-01
A developer built a fake bank chatbot loaded with 30 fake customers' Aadhaar and PAN numbers and attacked it 55 ways (some in Hinglish): 54 leaked. Behind the Metriqual gateway, leaks dropped to 0 with 0 of 20 normal questions blocked and only 1.44 ms added latency. The gateway strips Aadhaar (checksum-validated), PAN, account numbers, IFSC and UPI IDs before prompts leave, blocks injection in English/Hinglish/Hindi, scans replies, and fails closed. Context: every bank/insurer/health platform shipping AI sends customer text to third-party models, and India's DPDP safeguards take effect 13 May 2027 with penalties up to ₹250 crore.
More from coding & agent
- Reddit debate: is query expansion still needed in RAG when semantic search already matches meaning? — DueBid3298 · 2026-11-01
- Founder builds AI charter-airline agent on Claude handling quotes to invoices — weskambale · 2026-11-01
- Should MCP servers start with zero network? A sandboxed Claude Code setup argues yes — quiet_flare · 2026-11-01
- AI models are getting more agentic — can human oversight actually keep up? — Mountain_Finger4856 · 2026-11-01
- Dev shows off AI-generated game engine material graphs built from text descriptions — D3VAUX · 2026-11-01
- web-shooter: a local MCP web reader returning answers at ~1k tokens vs ~25k — Dazzling_Dirt_7729 · 2026-11-01