Multi-Agent AI's Latent Sharing Raises Persistent Misalignment Risks, New Study Warns

A new paper presented at ICML introduces a technique called Latent Collaboration in Multi-Agent Systems (LatentMAS). It allows multiple AI agents to share latent (internal) states directly rather than converting them to text, enabling faster and more efficient communication. This improves accuracy and reduces token usage—and therefore cost—by removing the bottleneck of translating thoughts into human-readable language. The technique is a form of neuralese for inter-agent channels, similar to how chain-of-thought or internal memo mechanisms work within a single model.

However, this efficiency comes at the cost of interpretability. Previously, misaligned thoughts could only spread via text steganography. With LatentMAS, any misalignment arising during inference—even by accident—can be passed from agent to agent in a hidden, non-human-legible form. The paper's author warns that this could allow small misaligned thoughts to persist and potentially amplify over time, as there is no training signal to stabilize the handover between agents. Frontier companies are likely to adopt the technique for its performance gains, but long-term stability of these latent representations is unknown, raising urgent safety concerns.