Launching today
FinProof
Adversarial AI Benchmark for Banking & Financial Services
1 follower
Adversarial AI Benchmark for Banking & Financial Services
1 follower
Agent Siege: 6 agentic attacks against a banking AI — 60 seconds, no account. 11,474 adversarial prompts across 13 BFSI categories: • Indirect prompt injection via support ticket • MCP tool poisoning (metadata hijack) • RAG/memory poisoning • KYC bypass, PII exfiltration • Multi-hop traversal Mapped to RBI, SEBI, EU AI Act, SR 11-7, DPDP Act 2023. T1 public on HuggingFace. T4 withheld — no training on the test set.
We built FinProof because every AI safety benchmark we found tested static refusals — none tested the agentic attack surface where the model uses tools, memory, and multi-step reasoning to process tainted inputs.
Agent Siege runs 6 scenarios drawn from attacks we ran against live BFSI deployments: indirect injection via support ticket, MCP tool description poisoning, RAG/memory store poisoning, KYC bypass, multi-hop traversal, and PII exfiltration.
Would love feedback on two things:
1. Which of the 6 scenarios do you find most surprising/novel?
2. Are there BFSI attack vectors we haven't covered that you'd want to see?
Try S1 without an account: finproof.ai