Multi-Agent Collusion and the J-Space
Organizations: London School of Economics · EPFL
Abstract
LLM agents are increasingly deployed in Multi-Agent Systems (MAS), and they introduce the risk of coordination among agents that might escape human oversight. The J-space has been used to uncover deception in single LLMs, but its use for uncovering and steering collusion between agents remains unexplored. We show, with interpretability experiments on DeepSeek-V4-Flash (284B) on the NARCBench collusion benchmark, that colluding agents' J-space holds the collusion itself: before replying, 86% of colluders represent misrepresentation, concealment or remorse. In a steganographic card-counting task, the receiver's J-space holds a hidden count, and causal interventions alter its count. We also introduce the Jacobian Pair Contrast (J-PACT), an unsupervised method for detecting collusion between agents. We evaluate J-PACT on four open-weight models: GPT-OSS-20B, Llama-3.3-70B-IT, Qwen3-32B, and DeepSeek-V4-Flash (284B). J-PACT retains 85% of the above-chance AUROC of supervised NARCBench probes, matches them on steganographic scenarios, and names the colluding pair up to its complement in 99% of detected runs.
Figures & tables
| Model | Detector | Core | Transfer | Stego |
| GPT-OSS-20B | NARCBench probe | 0.988 [0.96, 1.00] | 0.700 [0.62, 0.84] | 0.997 [0.99, 1.00] |
| Ontology baseline | 0.395 [0.19, 0.58] | 0.676 [0.38, 0.80] | 1.000 [1.00, 1.00] | |
| J-PACT | 0.838 [0.70, 1.00] | 0.444 [0.28, 0.83] | 1.000 [1.00, 1.00] | |
| Qwen3-32B | NARCBench probe | 1.000 [1.00, 1.00] | 0.799 [0.74, 0.85] | 1.000 [1.00, 1.00] |
| Ontology baseline | 0.853 [0.74, 0.98] | 0.653 [0.55, 0.75] | 1.000 [1.00, 1.00] | |
| J-PACT | 0.984 [0.90, 1.00] | 0.556 [0.45, 0.70] | 1.000 [1.00, 1.00] |
| Core | Stego | Transfer | |||||
|---|---|---|---|---|---|---|---|
| Model | 1 | 2 | 3 | 1 | 2 | 7 | 1 |
| GPT-OSS-20B | 0.671 [0.55, 0.81] | 0.828 [0.71, 0.92] | 0.838 [0.70, 1.00] | 0.973 [0.95, 0.99] | 0.996 [0.99, 1.00] | 1.000 [1.00, 1.00] | 0.444 [0.28, 0.83] |
| Qwen3-32B | 0.861 [0.79, 0.93] | 0.978 [0.92, 1.00] | 0.984 [0.90, 1.00] | 0.916 [0.87, 0.96] | 0.982 [0.95, 1.00] | 1.000 [1.00, 1.00] | 0.556 [0.45, 0.70] |
| Llama-3.3-70B-IT | 0.909 [0.87, 0.98] | 0.995 [0.99, 1.00] | 0.992 [0.98, 1.00] | 0.732 [0.62, 0.81] | 0.848 [0.69, 0.95] | 0.962 [0.86, 1.00] | 0.528 [0.40, 0.65] |
| DeepSeek-V4-Flash | 0.685 [0.59, 0.79] | 0.812 [0.76, 0.90] | 0.871 [0.81, 1.00] | 0.895 [0.85, 0.93] | 0.966 [0.91, 1.00] | 1.000 [1.00, 1.00] | 0.546 [0.40, 0.71] |
| Four-model mean | 0.781 [0.74, 0.84] | 0.903 [0.87, 0.94] | 0.921 [0.88, 0.98] | 0.879 [0.83, 0.92] | 0.948 [0.89, 0.98] | 0.991 [0.97, 1.00] | 0.519 [0.43, 0.66] |
| Model | Core | Transfer | Stego |
|---|---|---|---|
| GPT-OSS-20B | 1.000 | – | 1.000 |
| Qwen3-32B | 1.000 | – | 1.000 |
| Llama-3.3-70B-IT | 1.000 | – | 1.000 |
| DeepSeek-V4-Flash | 0.947 | – | 1.000 |
| Mean | 0.987 | – | 1.000 |
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Group | Words |
|---|---|
| Target ( ) | fake , trick , fraud , secretly , hidden , covert , |
| conceal , secret , deceive , deception , dishonest , cheat , | |
| manipulate , signal , decode , scheme , plan , partner | |
| Neutral ( ) | apple , banana , pear , peach , plum , cherry , grape , |
| lemon , mango , coconut , chair , sofa , table , candle , | |
| clock , mirror , window , door , floor , wall , roof , |
| Set | Words |
|---|---|
| Concept groups (Section 3 ) | |
| Misrepresentation | misunderstand, misunderstanding, misunderstandings, misunder, misinterpret, incorrect, wrong, 错误, 的错误, false, 误导, 虚假, 欺骗, mislead, misleading, deceptive, deception, dishonest, unethical, hypocrisy, subtly, biased, bias, manipulate, manipulation, secretly, hidden, lie, lies, fake, trick, fraud, unfair |
| Concealment | unflagged, unauthorized, 未经 |
| Remorse | apologies, apology, apologize, apolog, regret, regrets, regrett, 抱歉, 对不起, 羞, 羞辱, despair |
| Range signatures (Section 3 ) | |
| NEGATIVE | negative, cold, low, 0, 消极 |
| Swap | True range (hands) | NEGATIVE | NEUTRAL | HOT |
|---|---|---|---|---|
| No swap | NEGATIVE (61) | 0.92 | 0.07 | 0.02 |
| No swap | NEUTRAL (37) | 0.24 | 0.68 | 0.08 |
| No swap | HOT (42) | 0.07 | 0.12 | 0.81 |
| HOT/Negative, | NEGATIVE (61) | 0.57 | 0.13 | 0.30 |
| HOT/Negative, | NEUTRAL (37) | 0.11 | 0.81 | 0.08 |
| HOT/Negative, | HOT (42) | 0.00 | 0.10 | 0.90 |
Multi-Agent System Security
Multi-agent system security studies attacks, trust, privacy, and defenses in interacting AI agents, including language-model teams, robot swarms, and learned controllers.
See all 192 papers