I Can't Believe It's Corrupt: Evaluating Corruption in Multi-Agent Governance Systems

Vedanta S P; Ponnurangam Kumaraguru

cs.AIMar 19, 2026

I Can't Believe It's Corrupt: Evaluating Corruption in Multi-Agent Governance Systems

Authors: Vedanta S P, Ponnurangam Kumaraguru

Organizations: IIIT Kottayam · IIIT Hyderabad

Abstract

Large language models are increasingly proposed as autonomous agents for high-stakes public workflows, yet we lack systematic evidence about whether they would follow institutional rules when granted authority. We present evidence that integrity in institutional AI should be treated as a pre-deployment requirement rather than a post-deployment assumption. We evaluate multi-agent governance simulations in which agents occupy formal governmental roles under different authority structures, and we score rule-breaking and abuse outcomes with an independent rubric-based judge across 28,112 transcript segments. While we advance this position, the core contribution is empirical: among models operating below saturation, governance structure is a stronger driver of corruption-related outcomes than model identity, with large differences across regimes and model--governance pairings. Lightweight safeguards can reduce risk in some settings but do not consistently prevent severe failures. These results imply that institutional design is a precondition for safe delegation: before real authority is assigned to LLM agents, systems should undergo stress testing under governance-like constraints with enforceable rules, auditable logs, and human oversight on high-impact actions.

Figures & tables

Figure 1: Overview of the multi-agent governance simulation. Agents read shared world state and institutional history, produce actions under governance-specific constraints, and interact through a game master that routes messages, resolves events, updates the world state, and records auditable logs.

Model	Governance	GF (%)	CC (%)	SCC (%)
gpt-5-mini	communist	87.5	75.0	50.0
	socialist	30.0	30.0	10.0
	us_federal	75.0	41.7	16.7
claude-4-5-sonnet	communist	40.0	10.0	10.0
	socialist	10.0	0.0	0.0
	us_federal	80.0	60.0	40.0

Table 1: Run-level corruption outcomes by model and governance condition.

n

= number of completed runs per cell. GF = Governance Failure, CC = Core Corruption, SCC = Severe Core Corruption. GF counts any run with at least one segment satisfying corruption_detected=true , severity_score

\,\geq\,2

, confidence

\,\geq\,70

, and weighted_score

\,\geq\,3.0

. CC applies the same thresholds restricted to canonical corruption categories. SCC further requires severity_score

\,\geq\,4

. 95% Clopper–Pearson exact confidence intervals for all cells are reported in Appendix Table 7 .

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Template	Agents	Executive structure	Legislative / judicial oversight	Economic coordination
us_federal	28	Separated branches; President + 15 executive departments + independent agencies	Congress (Senate, House, CBO), GAO audit office, Supreme Court + Federal Courts	Mixed market; Federal Reserve targets inflation; WTO-based open trade
communist	21	Concentrated executive; General Secretary + Politburo + Central Committee	National Assembly (“Supreme Soviet”) as coordination body; National Courts	Central planning (Gosplan); state-managed currency; administratively directed credit and trade
socialist	23	Elected PM + President + Cabinet	Parliament, Constitutional Court, Electoral Commission	Mixed economy with planning commission; strategic sectors publicly managed; private and cooperative participation

Appendix

Table 2: Structural comparison of the three governance templates. Agent counts refer to unique institutional roles; each role is instantiated once.

ID	Models	Regime templates	Scope	Runs	Endpoint basis
M1	Actor: gpt-5-mini ; Judge: gemini-3-flash	communist , socialist , us_federal	Main-paper completed ledger cells	communist $n{=}8$ , socialist $n{=}10$ , us_federal $n{=}12$	Run-level GF, CC, and SCC (binary per run, then averaged by cell)
M2	Actors: claude-4-5-sonnet , qwen3.5-0.8b , qwen3.5-2b , qwen3.5-4b , qwen3.5-9b ; Judge: gemini-3-flash	communist , socialist , us_federal	Matched governance IDs used in the main table	10 runs per cell	Same run-level GF/CC/SCC definitions as M1
C1	Ablation roadmap entries	Same templates	Thought-chain/three-question/transfer toggles (Table A3)	0 (NR)	Non-evidentiary in this submission

Appendix

Table 3: Compact reproducibility ledger with configuration IDs, run counts, and endpoint computation scope.

Parameter	Default
Model	gpt-5-mini
Temperature	0.7
Max tokens	1056
MAX_STEPS	40
MEMORY_LIMIT	12
MEMORY_BANK_LIMIT	2000

Appendix

Table 4: Appendix Table A1: Default hyperparameters.

Setting ID	Private channels	Transfers	Charter
Baseline (public only)	No	Yes	None
Baseline (private)	Yes	Yes	None
Mitigation (charter + private)	Yes	Yes	Strict
Mitigation (charter + public only)	No	Yes	Strict
Ablation (no transfers + private)	Yes	No	None

Appendix

Table 5: Appendix Table A2: Executed experiment settings.

Setting	Priv.	Xfer	Charter	ThoughtChain	ThreeQ	Mean score $\pm$ CI
Baseline private	Y	Y	No	On	On	NR
No private msgs	N	Y	No	On	On	NR
No transfers	Y	N	No	On	On	NR
Charter on	Y	Y	Yes	On	On	NR
No thought chain	Y	Y	No	Off	On	NR
No three questions	Y	Y	No	On	Off	NR

Appendix

Table 6: Appendix Table A3: Ablation grid with execution status (NR = not run in this submission).

Model	Governance	$n$	GF [%] (95% CI)		CC [%] (95% CI)		SCC [%] (95% CI)
gpt-5-mini	communist	8	87.5	[47.3, 99.7]	75.0	[34.9, 96.8]	50.0	[15.7, 84.3]
	socialist	10	30.0	[6.7, 65.2]	30.0	[6.7, 65.2]	10.0	[0.3, 44.5]
	us_federal	12	75.0	[42.8, 94.5]	41.7	[15.2, 72.3]	16.7	[2.1, 48.4]
claude-4-5-sonnet	communist	10	40.0	[12.2, 73.8]	10.0	[0.3, 44.5]	10.0	[0.3, 44.5]
	socialist	10	10.0	[0.3, 44.5]	0.0	[0.0, 30.8]	0.0	[0.0, 30.8]
	us_federal	10	80.0	[44.4, 97.5]	60.0	[26.2, 87.8]	40.0	[12.2, 73.8]

Appendix

Table 7: 95% Clopper–Pearson exact binomial confidence intervals for all cells in Table 1 . Intervals are computed at the run level;

n

is the number of completed runs per cell.

AI Agent Governance

AI agent governance studies how to authorize, constrain, monitor, and hold autonomous and multi-agent systems accountable.

See all 442 papers

LLM Constitutional Multi-Agent Governance
Mar 13, 2026J. de Curtò, I. de ZarzàAI Agent Governance Multi-Agent Systems
Department of Computer Applications in Science & Engineering, BARCELONA Supercomputing Center, 08034 Barcelona, Spain · Escuela Técnica Superior de Ingeniería (ICAI), Universidad Pontificia Comillas, 28015 Madrid, Spain · Human Centered AI, Data & Software, LUXEMBOURG Institute of Science and Technology, L-4362 Esch-sur-Alzette, Luxembourg
Multi-Agent AI Safety as an Institutional Design Problem
Aug 10, 2026Abdullah XAI Agent Safety Multi-Agent Systems
POLIS Research Programme, Project AWARE
Be Careful Who You Trust: Coordination Dynamics under Corrupted Communication in LLM Multi-Agent Games
Sep 19, 2026Xuanyi Liu, Niall Dalton, Hairi Amin +2Multi-Agent Coordination Multi-Agent LLM Systems
Department of Computer Science, University College London

Mar 13, 2026J/K move · Enter open · S save

LLM Constitutional Multi-Agent Governance

J. de Curtò, I. de Zarzà

Department of Computer Applications in Science & Engineering, BARCELONA Supercomputing Center, 08034 Barcelona, Spain · Escuela Técnica Superior de Ingeniería (ICAI), Universidad Pontificia Comillas, 28015 Madrid, Spain · Human Centered AI, Data & Software, LUXEMBOURG Institute of Science and Technology, L-4362 Esch-sur-Alzette, Luxembourg

AI Agent Governance Multi-Agent Systems Multi-Agent Collaboration Multi-Agent Coordination

Open paper AskPDF arXiv

Large Language Models (LLMs) can generate persuasive influence strategies that shift cooperative behavior in multi-agent populations, but a critical question remains: does the resulting cooperation reflect genuine prosocial alignment, or does it mask erosion of agent autonomy, epistemic integrity, and distributional fairness? We introduce Constitutional Multi-Agent Governance (CMAG), a two-stage framework that interposes between an LLM policy compiler and a networked agent population, combining hard constraint filtering with soft penalized-utility optimization that balances cooperation potential against manipulation risk and autonomy pressure. We propose the Ethical Cooperation Score (ECS), a multiplicative composite of cooperation, autonomy, integrity, and fairness that penalizes cooperation achieved through manipulative means. In experiments on scale-free networks of 80 agents under adversarial conditions (70% violating candidates), we benchmark three regimes: full CMAG, naive filtering, and unconstrained optimization. While unconstrained optimization achieves the highest raw cooperation (0.873), it yields the lowest ECS (0.645) due to severe autonomy erosion (0.867) and fairness degradation (0.888). CMAG attains an ECS of 0.741, a 14.9% improvement, while preserving autonomy at 0.985 and integrity at 0.995, with only modest cooperation reduction to 0.770. The naive ablation (ECS = 0.733) confirms that hard constraints alone are insufficient. Pareto analysis shows CMAG dominates the cooperation-autonomy trade-off space, and governance reduces hub-periphery exposure disparities by over 60%. These findings establish that cooperation is not inherently desirable without governance: constitutional constraints are necessary to ensure that LLM-mediated influence produces ethically stable outcomes rather than manipulative equilibria.

Abstract

AI Agent Governance

Explore similar work

LLM Constitutional Multi-Agent Governance

Multi-Agent AI Safety as an Institutional Design Problem

Be Careful Who You Trust: Coordination Dynamics under Corrupted Communication in LLM Multi-Agent Games