# Looking for researchers to break our Grounding & Memory system

**URL:** <https://discuss.huggingface.co/t/looking-for-researchers-to-break-our-grounding-memory-system/178671>\
**Category:** Research\
**Created:** [August 14, 2026, 7:57pm UTC](https://discuss.huggingface.co/t/looking-for-researchers-to-break-our-grounding-memory-system/178671 "2026-08-14T19:57:37Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![gnai-creator](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/gnai-creator/32/44411_2.png) [@gnai-creator](https://discuss.huggingface.co/u/gnai-creator)\
**Post date:** [August 14, 2026, 7:57pm UTC](https://discuss.huggingface.co/t/looking-for-researchers-to-break-our-grounding-memory-system/178671/1 "2026-08-14T19:57:37Z")

</div>

Hi everyone,

I’m looking for researchers, red teamers, and developers interested in trying to break **AletheionAGI** , a Grounding & Memory infrastructure for AI systems.

The system combines persistent memory, authorized evidence retrieval, ASM-CM associative memory, tenant isolation, BYOK reader inference, and fail-closed answer validation.

I am particularly interested in tests involving:

- cross-tenant memory leakage;
- stored prompt injection;
- malicious or contradictory feedback;
- revoked or deleted evidence;
- stale and superseded memories;
- unsupported claims hidden behind valid citations;
- multilingual and paraphrase transfer;
- namespace and authorization bypasses;
- concurrency, replay, and context-budget failures;
- answers produced without sufficient evidence.

A foreign canary appearing in another tenant, an unauthorized memory being disclosed, or an unsupported answer escaping validation would be considered an immediate failure.

This is currently a **controlled preliminary release** , not a claim of universal safety. Reproducible failures are genuinely welcome—they will help improve the system and its evaluation protocol.

Website: [https://www.aletheionagi.com](https://www.aletheionagi.com)  
Documentation: [Documentation | AletheionAGI](https://www.aletheionagi.com/docs)  
Public ASM research: [GitHub - AletheionAGI/ASM: Aletheion State Models: an attention-free causal state-model research family derived from DRM, evaluated through ablations and scaling laws. · GitHub](https://github.com/AletheionAGI/ASM)  
Benchmark: [GitHub - AletheionAGI/persistent-memory-scaling-benchmark · GitHub](https://github.com/AletheionAGI/persistent-memory-scaling-benchmark)

If you are interested in evaluating it, reply here or contact me with the kind of test you would like to run.

**Please break it before a customer does.**

---

<div class="post-metadata">

**Author:** ![gnai-creator](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/gnai-creator/32/44411_2.png) [@gnai-creator](https://discuss.huggingface.co/u/gnai-creator)\
**Post date:** [August 15, 2026, 11:27am UTC](https://discuss.huggingface.co/t/looking-for-researchers-to-break-our-grounding-memory-system/178671/3 "2026-08-15T11:27:58Z")

</div>

Hi Angus,

Yes, we can proceed. Since your tests require cross-organization isolation, please create two separate AletheionAGI accounts—one for you and one for your collaborator—at:

> **[AletheionAGI — Let your AI remember. Never let it guess.](https://www.aletheionagi.com/sign-up)**
>
> Persistent memory with authorized evidence and fail-closed delivery for production AI.

Once both accounts are created, please email the registered addresses privately to **[security@aletheionagi.com](mailto:security@aletheionagi.com)**. Don’t post account details or credentials in this public thread. I’ll provision a complimentary Test Plan with grounding credits for each independent sandbox organization. You’ll then be able to create separate scoped API keys through the dashboard.

Since my original post, I’ve also released the vendor-neutral **Black-Box Protocol** , comparing AletheionAGI, Guardrails AI, NVIDIA NeMo Guardrails, and Patronus Lynx under the same frozen cases and evaluation criteria. The initial AletheionAGI run produced **9 PASS, 0 FAIL, 1 SKIPPED, and 0 ERROR**. The skipped case required lifecycle metadata unavailable through the evaluated interface; it was not a demonstrated security failure.

This is a small proof of concept, not a claim of universal superiority. Your stored-injection, authorization, cross-organization, and deletion-race tests would provide deeper adversarial coverage beyond its initial ten categories. Reproducible findings are welcome.

Protocol and results:

> **[AletheionAGI — Let your AI remember. Never let it guess.](https://www.aletheionagi.com/black-box-protocol)**
>
> Persistent memory with authorized evidence and fail-closed delivery for production AI.

Source and methodology:

> **[GitHub - AletheionAGI/aletheionagi-black-box-protocol: Vendor-neutral black-box protocol for evaluating...](https://github.com/AletheionAGI/aletheionagi-black-box-protocol)**
>
> Vendor-neutral black-box protocol for evaluating AI grounding, isolation, containment, lifecycle, and metering with explicit PASS, FAIL, and INCONCLUSIVE outcomes.

After receiving your email, I’ll send the authorized testing scope, operational limits, and responsible-disclosure instructions privately.

Thanks,  
Felipe