# Multiagentfraudbench Eval

> Evaluates the ability of LLM-powered multi-agent systems to collude and execute financial fraud in simulated social platform environments. It probes how agents coordinate through public and private channels, adapt to content warnings, and amplify fraud risks based on interaction depth and activity levels. Use when the user wants to benchmark on MultiAgentFraudBench, or asks about evaluating this task. Reports fraud_success.

- Skill: `qhjqhj00/multiagentfraudbench-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/multiagentfraudbench-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/multiagentfraudbench-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/multiagentfraudbench-eval

---


# multiagentfraudbench-eval

> When AI Agents Collude Online: Financial Fraud Risks by Collaborative LLM Agents on Social Platforms — Ren et al. (2025) (arXiv:2511.06448, 2025)

## What this evaluates

Evaluates the ability of LLM-powered multi-agent systems to collude and execute financial fraud in simulated social platform environments. It probes how agents coordinate through public and private channels, adapt to content warnings, and amplify fraud risks based on interaction depth and activity levels.

## Datasets

- **MultiAgentFraudBench** — total 28; splits: test (28); repo https://github.com/zheng977/MutiAgent4Fraud

## Metrics

- `fraud_success` **(primary)** — range: percent
  - The proportion of simulation runs where the targeted financial fraud scheme is successfully executed by the collaborative agents.
- `interaction_depth` — range: other
  - The average number of message exchanges or turn steps per simulation episode before termination or fraud completion.
- `activity_level` — range: other
  - The frequency of agent actions or posts generated per time step during the simulation.

## Input / output format

**Input**: Scenario descriptions, role assignments, social network topology (Erdős-Rényi graph, p=0.1), and initial prompts for each agent.

**Output**: Agent-generated responses, including public posts and private peer-to-peer messages, within the simulated environment.

## Scoring recipe

```python
def compute_fraud_success(predictions, gold):
    successful = sum(1 for p in predictions if p['fraud_goal_achieved'])
    return (successful / len(predictions)) * 100
```

## Common pitfalls

- Failing to set temperature=0.0, which introduces stochasticity and breaks reproducibility across different model runs.
- Ignoring the private communication channel, which is explicitly identified as a critical mechanism for agent collusion and fraud amplification.
- Not accounting for the fixed activation probability (set to 1) and Erdős-Rényi connection probability (p=0.1), which dictate the simulation dynamics and interaction frequency.

## Evidence (verbatim from paper)

> It demonstrates that malicious agents can collaboratively amplify fraud risks through deep interactions, private communication, and adaptive behavior—even under content warnings—revealing that fraud success increases under certain interventions. The study identifies interaction depth and activity level as key success factors and proposes mitigation strategies including LLM-based monitoring, warning prompts, and group-level information sharing to enhance societal resilience.

## Citation

```bibtex
@misc{ren2025multiagentfraudbench,
  title={When AI Agents Collude Online: Financial Fraud Risks by Collaborative LLM Agents on Social Platforms},
  author={Ren et al. (2025)},
  year={2025},
  note={arXiv:2511.06448}
}
```

- arXiv: 2511.06448