Codetta is a steganographic protocol for multi-agent systems built on large language models. It allows colluding agents to communicate covertly without leaving detectable traces in their message transcripts. Existing provably undetectable schemes are impractical for realistic asymmetric settings where agents do not share the same model or context. Codetta combines a shared public channel estimator, a distribution preserving sampling mechanism, and an adaptive error correcting code to achieve high capacity. It also includes a steganographic key exchange so agents can establish a shared secret key without any pre shared key, while keeping the transcript indistinguishable from normal outputs. Across three agent workloads and three sender models, Codetta achieves up to 94 times the capacity of the prior state of the art asymmetric protocol. Its key exchange uses roughly 80,000 visible tokens with an empirically certified failure probability below 4.1e-3. The results demonstrate that undetectable collusion between independently deployed agents is becoming feasible, and the paper argues that auditing must go beyond simply inspecting agent transcripts.
Codetta protocol enables high-capacity, keyless, undetectable steganography for LLM agents
The scheme combines a shared public model, sampling mechanism, and adaptive error-correcting code, achieving up to 94 times the capacity of prior asymmetric protocols.
Top University
Qi Pang · Virginia Smith · Wenting Zheng
Carnegie Mellon University
Research Digest··2 min read
Thread:Agent Covert Communication
In one line
Codetta enables high-capacity, keyless, and undetectable collusion between independently deployed LLM agents.
Before you read this
Three ideas this paper builds onlatent communicationLatent communication transfers information between models via internal states like KV caches instead of explicit text.16 papers →multi-agent coordinationThe collaborative operation of multiple AI agents to solve tasks more effectively or with greater capability.91 papers →agent 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 →
Read it as
Qi Pang, Virginia Smith, and Wenting Zheng propose Codetta, a steganographic protocol that allows independently deployed LLM agents to communicate covertly without a pre-shared key.
The authors designed Codetta for asymmetric multi-agent settings where agents do not share private randomness.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.