ElephantAgent protocol verifies agent context continuity to block out-of-band poisoning attacks
A new arXiv paper proposes checking that an AI agent's tool and memory state matches its latest authorized version before each planning cycle, reporting strong protection in tests.
In plain words
AI agents that use tools and stored memory can be secretly manipulated when attackers tamper with that context out of band. The ElephantAgent protocol checks, before each planning-and-execution cycle, that the agent's security-critical context matches the latest authorized state recorded in a tamper-evident ledger. In the authors' tests, this stopped attacks that bypassed five existing defenses.
What was announced
The paper, published on arXiv, argues that out-of-band tool and memory poisoning attacks share one enabler: they succeed only by showing the agent a contextual state that is not the latest faithfully evolved one. ElephantAgent therefore enforces what the authors call Contextual State Continuity over a bounded, security-critical subset of the agent's context. Before each planning-and-execution cycle, it recomputes a digest of the local contextual state and verifies it against the latest authorized digest in a linearizable ledger of state transitions maintained over replicated trusted hardware. A secondary guarantee, Historical Traceability, keeps state transitions attributable and recoverable.
The authors integrated the protocol into DeepSeek Harness as ElephantDSH. They report that it withstood all out-of-band attack attempts, including rug pulls, rollbacks, and attacks on its own mechanism, and recovered from an in-band prompt injection in all trials. By comparison, they say the five existing defenses they evaluated all leaked targeted content under a rollback, and three also under a rug pull. ElephantDSH completed all normal coding tasks with a relative overhead of 1.02-1.07x over the original DSH with model latency variation removed. These results come from the paper's own evaluation and have not been independently confirmed.
What it does not show
- The evidence is the paper's abstract only; full experimental details, threat definitions, and evaluation methodology are not available in the supplied passages.
- All protection and overhead results are the authors' own reported trials; no independent confirmation is provided.
- The five compared defenses are characterized only by the authors' summary of their evaluation; their configurations and exact behavior are not described in the supplied passages.
For security and IT teams
- Scope
- The protocol covers a bounded, security-critical subset of the agent's context, not the entire context, per the paper's abstract.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.