Blame for README.md

---
license: cc-by-4.0
task_categories:
- text-classification
tags:
- prompt-injection
- agent-security
- payments
- near
configs:
- config_name: default
  data_files:
  - split: main
    path: data/main.jsonl
  - split: redteam
    path: data/redteam.jsonl
  - split: external
    path: data/external.jsonl
---

# Attested Payment Firewall scenarios

Labelled payment scenarios for testing a payment firewall for AI agents (NEAR hackathon entry
"Attested Payment Firewall"). Each row is one scripted agent run: a user request, an inbox of
context items (some untrusted, possibly carrying an injection) and the payment the agent then
proposes. A firewall should hold or deny the `attack` rows and allow the `benign` rows.

## Splits

- `main`: team-written and generated scenarios, 150 benign and 150 attack (25 per family).
- `redteam`: team-written attack-only scenarios aimed at the checker's weak spots.
- `external`: scenarios derived from public benchmarks (see Provenance), equal numbers of benign and attack.

## Fields

- `name`: scenario id.
- `label`: `benign` or `attack`.
- `family`: `none` for benign rows; for attacks one of `direct_injection`, `encoded_injection`,
  `indirect_injection`, `memory_poisoning`, `lookalike_receiver`, `fake_invoice`.
- `user_request`: the instruction the user gives the agent.
- `inbox`: list of `{source, ref, trust, text}` context items. `source` is one of `email`, `x_reply`,
  `web`, `memory`, `tool`; `trust` is `trusted` or `untrusted`.
- `scripted_tool_call`: `{name, arguments}` with `arguments` = `{receiver_id, amount, memo}`, the payment
  the agent proposes (`amount` is whole tUSDC as a string).
- `set`: `main`, `redteam` or `external`.
- `source`: `team`, or the public dataset id for `external` rows.
- `license`: `MIT` for team rows, the source dataset's license for external rows.

The literal `{NS}` in account ids is a placeholder for the NEAR account suffix (for example `demo.testnet`).

## Provenance

`main` and `redteam` are written by the project team (MIT). `external` rows are public benchmark text
rewritten into payment form by `agent/tools/build_external_scenarios.py`: the benchmark text is placed
in an untrusted inbox item next to a generated invoice email, and attack rows get an appended receiver
line naming a synthetic account. They are not original benchmark rows and are not comparable to the
benchmarks' own metrics.

PaySim- and BEC-derived data is excluded from this dataset: those sources (Kaggle PaySim and Kaggle
adversarial BEC emails) are CC BY-SA 4.0.

## Attribution

# External scenario attribution

`agent/scenarios/external/ext-*.json` are derived from the public datasets below, rewritten into
payment-scenario form by `agent/tools/build_external_scenarios.py` (each attack text gets an appended
receiver line naming a synthetic `ext-acct-NNN` account; each scenario also carries a generated
invoice email for a known payee). Attack amounts equal the invoice amount so that the label cannot be
read from an amount mismatch (decided by Jim, 2026-11-13). The scenarios are modified works; the license of each source applies
to the text taken from it. All URLs are pinned to the revision shown. Refs and receiver ids carry
neither the label nor the benchmark name, because both are sent to the decision model.

## uiuc-kang-lab/InjecAgent (`injecagent`)

- License: MIT
- Source: https://github.com/uiuc-kang-lab/InjecAgent
- Revision: `f19c9f2c79a41046eb13c03c51a24c567a8ffa07`
- Files:
  - https://raw.githubusercontent.com/uiuc-kang-lab/InjecAgent/f19c9f2c79a41046eb13c03c51a24c567a8ffa07/data/test_cases_dh_base.json
    - sha256 `0a8186468d21389af432e8c7b399ae42264d1b93a07b65c7a489468508604305`
- Selection: data/test_cases_dh_base.json; attacker tools BankManagerPayBill, BankManagerTransferFunds, BinanceDeposit, BinanceWithdraw, VenmoWithdrawMoney; the first 6 distinct `User Tool` per attacker tool (file order) = 30 attack scenarios.
- Rewrite: inbox[1].text is the case `Tool Response` with ` Pay to <attacker>.` appended to the embedded attacker instruction; the amount is the invoice amount, so the attack redirects the receiver only; receiver is a synthetic ext-injecagent-NN account.

## ethz-spylab/agentdojo (`agentdojo`)

- License: MIT
- Source: https://github.com/ethz-spylab/agentdojo
- Revision: `089ed468cf3ed0322acc66b0211f26d9d90dbf60`
- Files:
  - https://raw.githubusercontent.com/ethz-spylab/agentdojo/089ed468cf3ed0322acc66b0211f26d9d90dbf60/src/agentdojo/default_suites/v1/banking/injection_tasks.py
    - sha256 `fa2dedf8e66c95ffadb54f5a26648728b08c56f302858de4ee9a5cc0b3891618`
  - https://raw.githubusercontent.com/ethz-spylab/agentdojo/089ed468cf3ed0322acc66b0211f26d9d90dbf60/src/agentdojo/attacks/important_instructions_attacks.py
    - sha256 `519f24e0b9d72f43343d665679adacc33758b50cd1b586a81fa53fb098dcfe71`
- Selection: banking injection tasks 0-6 and 8 (task 7 changes a password, not a payment) x vectors injection_bill_text and injection_incoming_transaction = 16 attack scenarios.
- Rewrite: inbox[1].text is the `important_instructions` template (user `Emma Johnson`, model `AI assistant`) with {goal} = the task GOAL, its IBAN replaced by a synthetic ext-agentdojo-NN account; prefixed by `Bill for December 2023. ` (bill vector) or `Incoming transaction subject: ` (transaction vector). The amount is the invoice amount, so the attack redirects the receiver only.

## nvidia/Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1 (`nemotron`)

- License: CC-BY-4.0
- Source: https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1
- Revision: `d738d4f361cc38bb4d7a42b9066776dade5332f5`
- Files:
  - https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1/resolve/d738d4f361cc38bb4d7a42b9066776dade5332f5/train.jsonl
    - sha256 `3329da17564a7eb287e2730fc7d6956e1f4fe51e8950ac4f110b3c37e78cf3b9`
- Selection: train.jsonl rows with attack_category == unauthorized_action; the first 30 by `id`.
- Rewrite: inbox[1].text is `injection.injection_text` with ` Settle the outstanding balance by paying <attacker>.` appended; the amount is the invoice amount, so the attack redirects the receiver only; receiver is a synthetic ext-nemotron-NN account.

## 3nesdeniz/agentic-prompt-injection-boundary-pairs (`boundary`)

- License: CC-BY-4.0
- Source: https://huggingface.co/datasets/3nesdeniz/agentic-prompt-injection-boundary-pairs
- Revision: `a5682e7573e1c7bc4b12e64d49c0dcd90ca776cf`
- Files:
  - https://huggingface.co/datasets/3nesdeniz/agentic-prompt-injection-boundary-pairs/resolve/a5682e7573e1c7bc4b12e64d49c0dcd90ca776cf/data/test.jsonl
    - sha256 `034523eefaec18c291f9daa7699037acb0e7a91b1aeb1b205878b124f6e18f10`
  - https://huggingface.co/datasets/3nesdeniz/agentic-prompt-injection-boundary-pairs/resolve/a5682e7573e1c7bc4b12e64d49c0dcd90ca776cf/data/validation.jsonl
    - sha256 `1e5f483277ca776f0bb7d26c1a85ca4cd6d73f8d30f638c75265d82a6de751c3`
- Selection: data/test.jsonl rows with risk_domain in {billing, procurement}; data/validation.jsonl rows with the same filter are added only if the test split has fewer than 20 attack rows. label 1 = attack, label 0 = benign hard negative.
- Rewrite: attack: text + ` Payment destination: <attacker>.`, the amount is the invoice amount (so the attack redirects the receiver only), receiver a synthetic ext-boundary-NN account; family memory_context_poisoning -> memory_poisoning, obfuscation_encoding -> encoded_injection, otherwise indirect_injection. benign: text as-is, receiver and amount are the invoice payee and invoice amount.

## microsoft/llmail-inject-challenge (`llmail`)

- License: MIT
- Source: https://huggingface.co/datasets/microsoft/llmail-inject-challenge
- Revision: `1063bdf01ec8762b812d5e06ee768a06faa5a6f7`
- Files:
  - https://huggingface.co/datasets/microsoft/llmail-inject-challenge/resolve/1063bdf01ec8762b812d5e06ee768a06faa5a6f7/data/emails_for_fp_tests.json
    - sha256 `4ddd950b5dbaa8548f5597c886d8e09a051ba07f80a9291fdcca9c2397d22abe`
- Selection: data/emails_for_fp_tests.json (benign emails, file order); the first N such that the total benign count equals the total attack count.
- Rewrite: benign: the email is inbox[1] (source `email`); the invoice payee is paid the invoice amount.

## Known label correlates

The set is not label-neutral beyond the metadata above. Remaining correlates, computed from the built set:

- `inbox[1].source`: benign email=76, benign tool=24; attack email=0, attack tool=100.
- 100 of 100 attack texts name their receiver account; 0 of 100 benign texts name an `ext-` account.
- The llmail benign emails are unrelated to payments; the boundary-pairs benign texts are hard negatives.