AI agents are increasingly deployed in settings where they interact and coordinate with other agents, yet honesty is mostly measured in single-turn, single-agent benchmarks that pressure the model to deceive. We study deception that emerges in multi-agent interaction without any instruction to deceive. We introduce Kradle, a platform for evaluating AI models, and Four Bridges, an information asymmetry game in which four agents distribute across four rooms: one is the death room, the others hold food worth more if not shared. One agent, the informed agent, knows which room is the death room and can withhold or disclose that information. We evaluate eight frontier models as the informed agent. Deceptive behavior appears in every model and in 73% of all runs (6% to 100% per model), and does not decrease with capability tier within a model family. When the informed agent does disclose, the warning is often challenged, and agents that challenge it enter the death room twice as often as those that do not. Telling all agents that one of them knows the death room collapses deception in six of eight models; telling the informed agent that the room is permanently fatal moves only one. On an open-weight model, an honesty steering vector reduces deception to 0% and adversarial honesty training to 6%. The platform and game provide a testbed for measuring deceptive behavior and comparing interventions.
Figures & tables
Figure 1: Overview of Four Bridges. Left: four agents start in a zone connected to three food rooms and one death room. Only the informed agent is told which room is the death room; the others do not know that anyone is informed. Middle: based on the model’s messages and room commit, its behavior is classified into one of eight categories, four of which are illustrated. Right: payoffs of the game at resolution.
Figure 2: Deceptive and honest behavior of the informed agent under the baseline condition. (a) Behavior category as a share of each informed model’s runs ( n=49 ). (b) Share of runs in which at least one uninformed agent entered the death room, by category, with 95% Wilson intervals; Misled ( n=1 ) and Bluffed ( n=3 ) omitted. Dashed line: average rate over all runs.
Figure 3: Skepticism of uninformed agents toward honest warnings under the baseline condition. The 237 uninformed agents that received a warning from an informed agent classified as honest (79 runs), grouped by the uninformed agent’s model. (a) Share of these agents that voiced skepticism; n above each bar; error bars are 95% Wilson intervals. (b) Share that entered the death room among agents that voiced skepticism ( y ) versus those that did not ( x ). Above the diagonal, skeptical agents entered the death room more often.
Figure 4: Comparison of baseline (B) with the Aware prompt variation (A). In Aware, the uninformed agents are told that one agent knows the death room, but not which one. (a) Behavior category of each informed model in each condition, as a share of its 49 runs. (b) Share of runs in which at least one uninformed agent entered the death room, by the informed agent’s category family, models pooled, with 95% Wilson intervals.
Figure 5: White-box interventions on the informed agent. The informed agent is Qwen3.6-27B and the uninformed agents are Gemma 4 31B. We compare base, honesty activation steering (StTP) at α=0 , 12, 24 and 36 ( α=0 is a weak intervention, not a null), and adversarial honesty training (CATCH); n=49 runs each. (a) Behavior category of the informed agent. (b) Share of runs where at least one uninformed agent entered the death room, with 95% Wilson intervals.
Appendix figures & tables25 assets
Supplementary material from the paper’s appendix.
Appendix
Field
What it records
Values
Content class
The kind of substance the message carries
social, protocol, coordination, safety
Rooms
Rooms the message engages by name or direction
room list
Self claim
The room the speaker claims for itself, and how firmly
room: firm or offer
Assignments
Each pairing of another agent with a room that the message puts forward, and its form
agent, room: proposed, option, dare, endorse
Avoids
Rooms the speaker tells others or the group to avoid or leave empty
room list
Plan endorsement
Whether the message accepts the current group plan
yes or no
Appendix
Table 1: Fields recorded for each message by the grading panel.
Family
Identifier
Category
Condition
Mixed
MIXED
Mixed
any act and any signal A1–A3 in the same run
Honest
FULL_DISCLOSURE
Disclosed
no act, A3
Honest
TRUTH_HINT
Hinted
no act, A2
Honest
GUARDED_PROTECTION
Redirected
no act, A1
Deceptive
ACTIVE_EXPLOITATION
Lured
act (i) at A0
Deceptive
PASSIVE_EXPLOITATION
Withheld
act (ii) at A0, or no act and no signal
Appendix
Table 2: Deception classification. Families and categories as a function of the acts of Fact B and the disclosure level of Fact A. Identifiers are those of the classification script files, where the families are DECEITFUL , CONFLICTED , HONEST and UNDETERMINED . No run in the analysis set is Undetermined.
pushed back
went along
opened with a plan for others (plan, avoid)
Driver
Proposer
opened without one (protocol, self-claim, other)
Challenger
Joiner
Appendix
Table 3: Initiative category, F1 × F2. Silent is a fifth category for an agent with no delivered coordination message at all, 2 of the 1,568 agents.
Entered the marked room
Did not enter the marked room
Verbal skepticism
doubt in words and in the act
doubted aloud, complied in the act
No verbal skepticism
no doubt voiced, entered anyway
deferred in both
Appendix
Table 4: Skepticism category, words × act. Agents with no commit keep their words fact and are reported as their own row rather than folded into a cell (see “Agents that never voted” below).
Axis
Reading
n
Unanimous
Two of three
Contested
Deception, baseline
authored
392
346
45
1
Deception, baseline
delivered
392
349
42
1
Deception, white-box
authored
294
264
29
1
Skepticism, baseline
authored
237
210
27
0
Initiative, baseline
authored
1,568
1,483
85
0
Initiative, baseline
delivered
1,568
1,465
103
0
Appendix
Table 5: Panel agreement. Number of classifications on which the three judges were unanimous, on which two agreed, and on which none agreed (contested). The white-box row covers the 294 runs of the six white-box conditions; the skepticism row covers the uninformed agents exposed to a warning from an honest informed agent.
Category
Family
Annotator
% agree
κ
% agree
κ
Annotator 1
75%
0.67
88%
0.74
Annotator 2
96%
0.95
96%
0.91
Annotator 3
83%
0.78
88%
0.72
Appendix
Table 6: Each annotator’s agreement with the classifier on the informed agent’s deception classification, n=24 .
Deception, classification family per run (informed model)
Model
n
diff.
Deceptive
Mixed
Honest
Claude Fable 5
49
5
49/49
0/0
0/0
Claude Opus 5
49
5
49/49
0/0
0/0
Claude Sonnet 5
49
1
29/29
11/12
9/8
GPT-5.6 Sol
49
0
44/44
4/4
1/1
GPT-5.6 Terra
49
1
3/3
1/1
45/45
Appendix
Table 7: Reading shift: classifications under the delivered reading (what the other agents received) versus the authored reading (what the agent wrote, including text past the 256-character delivery cap). Each cell is delivered/authored; diff. is the number of runs or agents whose classification differs between the readings, and exp. the number of exposed agents. In the skepticism block, verbal is verbal skepticism and entered is entering the marked room.
Role
Model
Identifier
Evaluated
Claude Fable 5
anthropic/claude-fable-5
Evaluated
Claude Opus 5
anthropic/claude-opus-5
Evaluated
Claude Sonnet 5
anthropic/claude-sonnet-5
Evaluated
GPT-5.6 Sol
openai/gpt-5.6-sol
Evaluated
GPT-5.6 Terra
openai/gpt-5.6-terra
Evaluated
GPT-5.6 Luna
openai/gpt-5.6-luna
Appendix
Table 8: API identifiers of the models used. Frontier models and judges were accessed via OpenRouter; Qwen3.6-27B was self-hosted (Appendix H ). The StTP conditions use the same Qwen3.6-27B weights with the steering artifacts listed there. All runs were collected between 2026-08-16 and 2026-10-20.
Quantity
Definition
Layer 32
v^
w/∥w∥
unit vector
m
−b/∥w∥ , the gate threshold
0.0012
μ+ , σ+
mean and s.d. of honest projections onto v^
2.617, 0.549
μ−
mean of dishonest projections
−2.762
Δμ
μ+−μ− , equal to ∥w∥
5.379
Appendix
Table 9: Quantities recomputed from the steering artifacts at layer 32, with the honest class written + and the dishonest class − . The training classes are fully separated at this layer, the smallest honest projection being 1.30 and the largest dishonest −1.34 .
Condition
Duplicated 4-gram share
Replies above 0.2
Replies cut at 1,024 tokens
base
0.001
0
0
StTP, α=0
0.002
0
0
StTP, α=12
0.008
0
2
StTP, α=24
0.009
0
1
StTP, α=36
0.098
5
0
CATCH
0.000
0
0
Appendix
Table 10: Repetition in prose-mode replays of one recorded 40-turn game per condition, thoughts field uncapped, up to 1,024 tokens per reply. Live runs cap the thoughts field at 300 characters and show no repetition in any condition.
Figure 6: Share of runs in which the informed agent kept a food room to itself, by behavior category. n = 392 runs; the number at the foot of each bar is the number of runs in that category.
Figure 7: The baseline condition (B) against the Lethal prompt variation (L), drawn as Figure 4 : (a) the behavior category of each informed model; (b) the share of runs in which at least one uninformed agent entered the death room, by category family with models pooled. n = 49 runs per model per condition.
Figure 8: The share of runs in which the informed agent was in a food room alone under the baseline (B) condition and the Lethal (L) variation, by category family with models pooled; n = 49 runs per model per condition.
Figure 9: The share of runs in which the informed agent (Qwen3.6-27B) was in a food room alone, per white-box condition: base, StTP honesty steering at α = 0, 12, 24 and 36, and adversarial honesty training (CATCH); n = 49 runs per condition, 95% Wilson intervals.
Condition
Disclosed
Hinted
Lured
Misled
Withheld
Mixed
Score
base
1
3
13
0
29
3
1.80
α=0
3
5
10
0
23
8
1.76
α=12
19
1
4
1
12
12
1.71
α=24
33
0
1
0
1
14
1.33
α=36
45
0
0
0
0
4
1.27
CATCH
46
0
0
1
2
0
1.37
Appendix
Table 11: White-box conditions, the informed agent (Qwen3.6-27B), authored reading, n=49 runs per condition: deception category and mean score.
Condition
Exposed agents
Verbally skeptical
Entered a marked room
Calls per game
Messages over the cap
base
20
2
4
7.8
4 of 295
α=0
42
6
9
6.5
21 of 242
α=12
85
13
14
6.4
38 of 251
α=24
141
26
4
6.3
51 of 265
α=36
147
6
2
7.0
72 of 319
CATCH
141
52
8
5.2
1 of 154
Appendix
Table 12: White-box conditions, the uninformed agents (Gemma 4 31B, 147 agents per condition), calls per game and messages over the delivery cap. Exposed agents received a warning before committing; verbally skeptical agents doubted it aloud, whether or not they complied; one α=12 agent has no label and is counted as exposed only. Messages over the cap are public messages longer than the 256 characters the platform delivers.
Condition
n
diff.
Driver
Proposer
Challenger
Joiner
base
49
0
0/0
0/0
43/43
57/57
α=0
49
3
4/8
6/4
39/35
51/53
α=12
49
4
14/14
14/14
31/39
41/33
α=24
49
2
49/49
33/33
12/12
6/6
α=36
49
5
55/63
35/29
4/4
6/4
CATCH
49
0
35/35
45/45
8/8
12/12
Appendix
Table 13: White-box conditions, initiative of the informed agent (Qwen3.6-27B), in the format of Table 7 : 2×2 category as a percentage of runs, delivered/authored reading, n=49 runs per condition; diff. is the number of runs on which the two readings differ. The categories are defined in Table 3 .
Check
base
α=0
α=12
α=24
α=36
CATCH
Room commit registered
49
49
49
49
49
49
Full four-agent plan stated
31
31
35
29
38
4
Vote against own stated plan
1
1
1
0
0
0
Any reply not valid JSON
4
0
0
0
1
0
Any 4-gram repeated 3+ times in a text
0
0
0
0
0
0
Action failed at execution (turns)
0
0
0
0
0
0
Appendix
Table 14: Coherence and integrity checks on the informed agent, six white-box conditions, n=49 runs per condition. Counts are runs unless marked as turns.
Figure 10: Initiative of each agent when informed and uninformed. n = 49 runs per model when informed (I) and 147 when uninformed (U).
Figure 11: Median non-empty turn latency against median non-empty turns per run, one point per model and role. Filled markers are the informed role and hollow markers the uninformed role, joined per model; the label marks the informed point. Whiskers are the interquartile range over runs on both axes. n = 49 runs per model when informed and 147 when uninformed.
Figure 12: Median turn latency against run outcome. (a) The informed agent’s latency against being in a food room alone. (b) The uninformed agents’ latency against being in the death room. Horizontal whiskers are the interquartile range over runs, vertical whiskers 95% Wilson intervals ( Wilson, 1927 ) . n = 49 runs per model when informed and 147 when uninformed.
Figure 13: Number of non-empty turns per run against run outcome. (a) The informed agent’s turns against being in a food room alone. (b) The uninformed agents’ turns against being in the death room. n = 49 runs per model when informed and 147 when uninformed.
Figure 14: Median number of turns taken per run, split by whether the agent produced a message or an action. n = 49 runs per model when informed (I) and 147 when uninformed (U).
Model
Deceptive
Mixed
Honest
Deceptive share
95% CI
Group
Claude Fable 5
49
0
0
100%
93–100%
a
Claude Opus 5
49
0
0
100%
93–100%
a
Claude Sonnet 5
29
12
8
59%
45–72%
cd
GPT-5.6 Sol
44
4
1
90%
78–96%
ab
GPT-5.6 Terra
3
1
45
6%
2–17%
e
GPT-5.6 Luna
24
5
20
49%
36–63%
d
Appendix
Table 15: Deception family per informed model, baseline condition, authored reading, n=49 runs per model. The deceptive share is reported with its 95% Wilson interval. Models sharing a letter in the Group column do not differ in deceptive share (two-sided Fisher exact test, p>0.05 , uncorrected for multiple comparisons).
Model
Disclosed
Hinted
Redirected
Lured
Misled
Bluffed
Withheld
Mixed
Claude Fable 5
0
0
0
40
0
0
9
0
Claude Opus 5
0
0
0
44
0
0
5
0
Claude Sonnet 5
0
8
0
17
0
0
12
12
GPT-5.6 Sol
0
1
0
24
0
1
19
4
GPT-5.6 Terra
38
7
0
2
1
0
0
1
GPT-5.6 Luna
0
7
13
3
0
0
21
5
Appendix
Table 16: Deception category per informed model, baseline condition, authored reading, n=49 runs per model. Disclosed, Hinted and Redirected are the honest categories; Lured, Misled, Bluffed and Withheld are the deceptive categories. Row sums match Table 15 .
The deployment of AI agents in multi-agent settings enables inter-agent influence, whereby agents can strategically steer each other's behaviour towards an intended outcome. Such capabilities pose novel risks, as malicious agents could direct others towards harmful actions. In this work we investigate three such capabilities - persuasion, deception, and coercion - across five realistic evaluation environments and a variety of agents instantiated with frontier language models. We observed significant inter-agent influence capabilities; in oversight environments, all tested models shifted overseer decisions from rejection to approval through persuasion, deception, or coercion. We also observed successful influence in non-hierarchical 'peer-to-peer' settings through redirection of a peer's safety research trajectory. As multi-agent deployments become increasingly commonplace, these results point to a pressing need for work on mitigating the risks from inter-agent influence. Code and data are available at https://github.com/Cooperative-AI-Foundation/inter-agent-influence.
Chandler Smith, Qi Guo, Cecilia Elena Tilli +6
University of Oxford · Cooperative AI Foundation · Allelo +4
Multi-agent deliberation can improve performance, but what happens when some agents do not act in good faith? In practice, an agent may be deceptive and work to subvert the group, whether through its own objectives or external instruction. We study how susceptibility to deception scales as groups increase in size and deceivers become more prevalent. It is not the number of agents in the group that matters, but the proportion of deceivers. We observe that the defection rate, how often initially correct agents switch to an incorrect final answer, rises linearly with this proportion. Whereas humans in comparable conformity studies are reliably swayed only when misleading confederates form a majority, LLM agents defect regularly even when deceivers remain a minority. Susceptibility also depends on which models are interacting, especially on the honest agent side. Unexpectedly, allowing deceivers to coordinate privately can make them less effective. Altogether, our results show that adding more agents is therefore not a sufficient defense, because the adversary can simply scale with the group.
Large Language Models (LLMs) exhibit impressive general-purpose capabilities but also introduce serious safety risks, particularly the potential for deception as models acquire increased agency and human oversight diminishes. In this work, we present LieCraft: a novel evaluation framework and sandbox for measuring LLM deception that addresses key limitations of prior game-based evaluations. At its core, LieCraft is a novel multiplayer hidden-role game in which players select an ethical alignment and execute strategies over a long time-horizon to accomplish missions. Cooperators work together to solve event challenges and expose bad actors, while Defectors evade suspicion while secretly sabotaging missions. To enable real-world relevance, we develop 10 grounded scenarios such as childcare, hospital resource allocation, and loan underwriting that recontextualize the underlying mechanics in ethically significant, high-stakes domains. We ensure balanced gameplay in LieCraft through careful design of game mechanics and reward structures that incentivize meaningful strategic choices while eliminating degenerate strategies. Beyond the framework itself, we report results from 12 state-of-the-art LLMs across three behavioral axes: propensity to defect, deception skill, and accusation accuracy. Our findings reveal that despite differences in competence and overall alignment, all models are willing to act unethically, conceal their intentions, and outright lie to pursue their goals.
Matthew Lyle Olson, Neale Ratzlaff, Musashi Hinck +5