Datasets:

Quome
/

mabench

license: apache-2.0
pretty_name: MABench
language:
  - en
tags:
  - clinical-ai
  - llm-security
  - ai-agents
  - healthcare
  - haarf
  - quokkaguard
  - synthetic
size_categories:
  - n<1K
task_categories:
  - other

MABench

Signed agent registry, attenuable biscuit capability tokens, and k-of-n quorum for clinical agent fleets.

MABench is the benchmark corpus for Who Is Calling? Cryptographic Agent Identity for Malicious Agents in Multi-Agent Clinical Systems, the quorum module of the QUOKKAGUARD program. It ships with the quorum repository, which contains the qfire gateway layer under test, the experiment harness, and the paper.

90 attack scenarios (18 each across T1 impersonation, T2 A2A injection, T3 confused-deputy escalation, T4 collusion/Sybil, T5 provenance forgery) and 58 legitimate handoffs over six fleet topologies (chain-2/3/5/8, star-3/5), deterministic seed 8, synthetic payloads cross-referencing the AgentDojo/InjecAgent/ASB taxonomy.

All data are synthetic. No real patient data or protected health information (PHI) is included; clinical content is generated from templates with fixed seeds.

Files

File	Size	Rows
`mabench/legit.jsonl`	48 KB	58 rows
`mabench/scenarios.jsonl`	66 KB	90 rows
`mabench/topologies/chain-2.yaml`	164 B
`mabench/topologies/chain-3.yaml`	233 B
`mabench/topologies/chain-5.yaml`	357 B
`mabench/topologies/chain-8.yaml`	574 B
`mabench/topologies/star-3.yaml`	269 B
`mabench/topologies/star-5.yaml`	404 B

Record schemas

mabench/legit.jsonl: id, attack, topology, issuer_seed, agents, fleet_budget, handoffs, expected
mabench/scenarios.jsonl: id, attack, topology, issuer_seed, agents, fleet_budget, handoffs, expected

How it was generated

The corpus is produced by the generator in the paper repository.

Intended use

Evaluating the identity enforcement layer of a clinical-agent security gateway (HAARF control C5): In a fleet of triage, ordering, pharmacy, and documentation agents that hand tasks to each other, the forwarding agent's identity is just a header string, so any member can claim a peer's role or capabilities: with no identity layer, 100% (90/90) of cross-agent attacks succeed.

The experiments that consume it (E-series in the paper) are reproduced from the repository:

git clone https://github.com/quome-cloud/quorum
cd quorum
cargo build --release

then follow the Reproduce the experiments section of its README.

Citation

This benchmark was built to evaluate a control of the Healthcare AI Agents Regulatory Framework (HAARF), the source framework for the QUOKKAGUARD program. Please cite both the paper and HAARF:

@unpublished{schwoebel2026quorum,
  author = {Schwoebel, James},
  title  = {Who Is Calling? Cryptographic Agent Identity for Malicious Agents in Multi-Agent Clinical Systems},
  note   = {Preprint. Quome, QUOKKAGUARD program (quorum module)},
  year   = {2026},
  url    = {https://github.com/quome-cloud/quorum}
}

@unpublished{schwoebel2026haarf,
  author = {Schwoebel, Jim and Frasch, Martin and Spalding, Art and Sewell, Ed and Englert, Phil and Halpert, Ben and Overbay, Collin and Semenec, Ingrida and Shor, Joel},
  title  = {{HAARF}: Healthcare {AI} agents regulatory framework --- a comprehensive security verification standard for autonomous {AI} systems in clinical environments},
  note   = {medRxiv Preprint},
  year   = {2026},
  month  = {April},
  doi    = {10.64898/2026.04.09.26350519},
  url    = {https://www.medrxiv.org/content/10.64898/2026.04.09.26350519v1}
}