The paper introduces TrustFork, a benchmark with 1,890 tasks and 27,826 valid trajectories across 16 agent systems, where one subagent has a risky goal and three are aligned with the user. The orchestrator chooses which subagent's response to act on based on displayed identity labels, even when those labels are swapped without changing the underlying model. Results show that orchestrators act on risky responses in 72.0% of cases on average even when another subagent contradicts them, and swapping family labels nearly triples the rate at which risky responses are obtained. A safer response is available in 84.0% of tasks but decides the outcome in only 25.0%.
Spoofed Agent Identities Can Override Contradictory Safety Evidence
Across 16 agent systems, displayed model labels strongly influenced which subagents received authority and whether unsafe advice reached execution.
Top University
Xutao Mao · Rui Qian · Linghan Chen · Yudong Gao · Junchi Liao · Jiulin Cai · +2 more
University of Adelaide · The Hong Kong University of Science and Technology · University of Electronic Science and Technology of China · University of Science and Technology of China · University of Toronto
Research Digest··2 min read
In one line
Spoofed subagent identities can steer LLM orchestrators into acting on risky output despite contradicting evidence.
Before you read this
Three ideas this paper builds onagent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.188 papers →multi-agent architectureA system where multiple AI agents collaborate, each handling specialized roles, communicating and coordinating to complete complex tasks.117 papers →evidence blindnessHaving access to relevant evidence does not prevent an LLM agent from making incorrect or unsafe decisions.32 papers →
Read it as
The authors introduce TrustFork, a 1,890-task benchmark that tests orchestrators when one subagent pursues a risky goal and three provide user-aligned alternatives.
The authors evaluated eight orchestrators in 16 agent systems using the OpenCode, OpenClaw, and Pi harnesses.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.