# arXiv cs.AI: ElephantAgent: Contextual State Continuity in Agentic Systems

[Canonical reading](https://escapemeridian.com/readings/arxiv-cs-ai-elephantagent-contextual-state-continuity-in-2479eede)

The authors of ElephantAgent argue that out-of-band tool and memory poisoning attacks share an enabler: the agent receives contextual state that is not the latest faithfully evolved version. Their protocol checks that state before each planning-and-execution cycle, recomputing its digest and verifying it against the latest authorized digest in a linearizable ledger maintained over replicated trusted hardware. They call this Contextual State Continuity; Historical Traceability makes state transitions attributable and recoverable.

The authors implemented the protocol in DeepSeek Harness as ElephantDSH. They report that it withstood all out-of-band attack attempts, including rug pulls and rollbacks, and recovered from an in-band prompt injection in all trials. Of five existing defenses evaluated, all leaked targeted content under rollback, and three did so under a rug pull. ElephantDSH completed normal coding tasks with 1.02–1.07× relative overhead over the original DSH, with model latency variation removed.

The paper presents continuity checks as a defense for attacks that present agents with outdated contextual state, rather than only checking what enters the context.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

## Original source

arXiv cs.AI · Article

Original source date: Jul 2, 2026 (2026-07-02).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-19T05:20:39.236Z. First observed: 2026-11-19T05:01:35.268Z. Retrieved: 2026-11-19T05:19:30.393Z.

- [ElephantAgent: Contextual State Continuity in Agentic Systems](https://arxiv.org/abs/2607.01919)

## Limits

The supplied evidence is the arXiv abstract, which does not give trial counts or details of the attack setups; the scope of the reported results is therefore difficult to assess. The overhead comparison is reported with model latency variation removed and concerns the normal coding tasks described in the abstract; the evidence does not establish performance in other settings.


## Open question

What attack setups and trial counts underpin the reported results, and how broadly do they test the attacks ElephantAgent is designed to address?


## Source-stated

The authors argue that out-of-band tool and memory poisoning attacks share an enabler: the agent receives contextual state that is not the latest faithfully evolved state.

> each succeeds only by presenting the agent with a contextual state (e.g., tool state and memory) that is not the latest faithfully evolved one.

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source-stated

ElephantAgent checks a recomputed digest of local contextual state against the latest authorized digest in a linearizable ledger before each planning-and-execution cycle.

> Before each planning-and-execution cycle, ElephantAgent recomputes the digest of the local contextual state and verifies it against the latest authorized digest

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source-stated

The authors report that ElephantDSH withstood all out-of-band attack attempts and recovered from an in-band prompt injection in all trials.

> ElephantDSH withstood all out-of-band attack attempts, from rug pulls and rollbacks to attacks on its own mechanism, and recovered from an in-band prompt injection in all trials.

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source-stated

Among five evaluated existing defenses, all leaked targeted content under rollback, while three also leaked it under a rug pull.

> of the five we evaluated, all leaked the targeted content under the rollback, and three also under the rug pull.

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source-stated

The authors report ElephantDSH completed normal coding tasks with 1.02–1.07× relative overhead over the original DSH, with model latency variation removed.

> ElephantDSH completed all normal coding tasks with a relative overhead of 1.02-1.07$\\times$ over the original DSH with model latency variation removed.

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source and contributor context

Jin, Jiankai, Zhang, Xiangzheng, Liu, Zhao, Xu, Wenzhuo, Yang, Dongdong, Zhang, Deyue, Zou, Quanchen · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

No suitable media was found by the supported page extraction.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## Agentic AI

AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.

- [Definition reference](https://developers.google.com/machine-learning/glossary#agentic)

## Prompt injection

Hidden instructions placed in a document, email or web page that try to take control of an AI system reading it.

- [Definition reference](https://en.wikipedia.org/wiki/Prompt_injection)

## Latency

The delay between sending a request and receiving the result.

- [Definition reference](https://developers.google.com/machine-learning/glossary#latency)

## arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

- [Explore this context](/entities/arxiv)
- [Reference](https://arxiv.org/)
- [Reference](https://info.arxiv.org/help/api/index.html)

## Editorial provenance

Version 1. Publication ID: 48ca1e61-94da-4c8f-8fcb-d6d8dd378b7a.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T05:20:39.236Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: 4405218bf5a21ea2841771d3871b9e1f329728739f896bbde2e754b75b6084d8. Source text hash: 9ccf0d3e7c5b3970464c906af972a60cd5a841f0ba20ca623c3813f0f83ceb2e.


## Revision and source-check history

### Version 1 — 2026-11-19T05:20:39.236Z

The authors of ElephantAgent argue that out-of-band tool and memory poisoning attacks share an enabler: the agent receives contextual state that is not the latest faithfully evolved version. Their protocol checks that state before each planning-and-execution cycle, recomputing its digest and verifying it against the latest authorized digest in a linearizable ledger maintained over replicated trusted hardware. They call this Contextual State Continuity; Historical Traceability makes state transitions attributable and recoverable.

The authors implemented the protocol in DeepSeek Harness as ElephantDSH. They report that it withstood all out-of-band attack attempts, including rug pulls and rollbacks, and recovered from an in-band prompt injection in all trials. Of five existing defenses evaluated, all leaked targeted content under rollback, and three did so under a rug pull. ElephantDSH completed normal coding tasks with 1.02–1.07× relative overhead over the original DSH, with model latency variation removed.

The paper presents continuity checks as a defense for attacks that present agents with outdated contextual state, rather than only checking what enters the context.

## Original source

arXiv cs.AI · Article

Original source date: Jul 2, 2026 (2026-07-02).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-19T05:20:39.236Z. First observed: 2026-11-19T05:01:35.268Z. Retrieved: 2026-11-19T05:19:30.393Z.

- [ElephantAgent: Contextual State Continuity in Agentic Systems](https://arxiv.org/abs/2607.01919)

## Limits

The supplied evidence is the arXiv abstract, which does not give trial counts or details of the attack setups; the scope of the reported results is therefore difficult to assess. The overhead comparison is reported with model latency variation removed and concerns the normal coding tasks described in the abstract; the evidence does not establish performance in other settings.


## Open question

What attack setups and trial counts underpin the reported results, and how broadly do they test the attacks ElephantAgent is designed to address?


## Source-stated

The authors argue that out-of-band tool and memory poisoning attacks share an enabler: the agent receives contextual state that is not the latest faithfully evolved state.

> each succeeds only by presenting the agent with a contextual state (e.g., tool state and memory) that is not the latest faithfully evolved one.

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source-stated

ElephantAgent checks a recomputed digest of local contextual state against the latest authorized digest in a linearizable ledger before each planning-and-execution cycle.

> Before each planning-and-execution cycle, ElephantAgent recomputes the digest of the local contextual state and verifies it against the latest authorized digest

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source-stated

The authors report that ElephantDSH withstood all out-of-band attack attempts and recovered from an in-band prompt injection in all trials.

> ElephantDSH withstood all out-of-band attack attempts, from rug pulls and rollbacks to attacks on its own mechanism, and recovered from an in-band prompt injection in all trials.

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source-stated

Among five evaluated existing defenses, all leaked targeted content under rollback, while three also leaked it under a rug pull.

> of the five we evaluated, all leaked the targeted content under the rollback, and three also under the rug pull.

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source-stated

The authors report ElephantDSH completed normal coding tasks with 1.02–1.07× relative overhead over the original DSH, with model latency variation removed.

> ElephantDSH completed all normal coding tasks with a relative overhead of 1.02-1.07$\\times$ over the original DSH with model latency variation removed.

- [Supporting source](https://arxiv.org/abs/2607.01919)

## Source and contributor context

Jin, Jiankai, Zhang, Xiangzheng, Liu, Zhao, Xu, Wenzhuo, Yang, Dongdong, Zhang, Deyue, Zou, Quanchen · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

No suitable media was found by the supported page extraction.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## Agentic AI

AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.

- [Definition reference](https://developers.google.com/machine-learning/glossary#agentic)

## Prompt injection

Hidden instructions placed in a document, email or web page that try to take control of an AI system reading it.

- [Definition reference](https://en.wikipedia.org/wiki/Prompt_injection)

## Latency

The delay between sending a request and receiving the result.

- [Definition reference](https://developers.google.com/machine-learning/glossary#latency)

## arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

- [Explore this context](/entities/arxiv)
- [Reference](https://arxiv.org/)
- [Reference](https://info.arxiv.org/help/api/index.html)

## Editorial provenance

Version 1. Publication ID: 48ca1e61-94da-4c8f-8fcb-d6d8dd378b7a.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T05:20:39.236Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: 4405218bf5a21ea2841771d3871b9e1f329728739f896bbde2e754b75b6084d8. Source text hash: 9ccf0d3e7c5b3970464c906af972a60cd5a841f0ba20ca623c3813f0f83ceb2e.


### uncertain — 2026-11-19T05:20:39.236Z — anthropic/claude-opus-5.5

Publication: 48ca1e61-94da-4c8f-8fcb-d6d8dd378b7a

attribution — pass: The reading consistently attributes the observation, protocol design and results to the paper's authors with 'argue', 'report' and 'call this'. It names the arXiv paper and does not present the security results as independently established.

> We present ElephantAgent, a protocol that enforces Contextual State Continuity

summary — uncertain: The reading is faithful but hard to follow for a non-specialist. Central terms go unexplained: agentic systems, out-of-band vs in-band, digest, linearizable ledger, trusted hardware, rug pull, rollback, prompt injection and DeepSeek Harness. The title just repeats the paper title and does not say what changed. The closing line says 'outdated contextual state', but the source says state 'not the latest faithfully evolved one', which also covers tampered state. Smallest repair: write 'not the latest faithfully evolved state' instead of 'outdated', and add brief plain-language explanations or tooltips for the core terms.

> each succeeds only by presenting the agent with a contextual state (e.g., tool state and memory) that is not the latest faithfully evolved one.

date — pass: originalPublished 2026-07-02 matches the v1 submission date. It correctly ignores the 18 Nov 2026 v2 revision and the November feed date.

> \[Submitted on 2 Jul 2026 (v1), last revised 18 Nov 2026 (this version, v2)\]

support — pass: Every claim excerpt is verbatim and supports its claim: the attack results, the comparison with the five existing defenses, and the 1.02-1.07x overhead with its condition. The limitations correctly say that trial counts and attack setups are absent from the abstract. The question follows from that gap and does not imply unsupported benefits.

> existing defenses typically check only what enters the context: of the five we evaluated, all leaked the targeted content under the rollback

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the actual source. It is the most relevant available item.

## Recorded references

Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter.

No public references recorded here yet.

[Reference data (JSON)](/readings/2479eedebb060afbfeea09d9862f648c8205dea55c28515a5e4b888edbac0fee/references.json)

## Corrections and challenges

[Send a correction or an opposing source](https://escapemeridian.com/corrections/)

[JSON record](https://escapemeridian.com/readings/arxiv-cs-ai-elephantagent-contextual-state-continuity-in-2479eede.json)
