Lab
A memory is data, not an instruction
Short answer
A memory can contain a decision, a quotation or a passage asking an assistant to change its behaviour. Those forms can look similar. For a memory intended to work across several AI systems, the challenge is to return useful information without granting it authority it does not possess.

When the material being read tries to direct the reader
Prompt injection attempts to make an assistant treat content as a higher-priority instruction. The risk is not limited to websites. A stored document can contain text designed to influence a later response.
Memories can also legitimately contain instructions. A team may document how it addresses customers, save a product prompt or quote a support example. Blocking every passage addressed to an AI would prevent many ordinary uses.
The research problem is therefore more demanding than detecting imperative language. It requires distinguishing text that reports an instruction, an instruction deliberately intended by the user and an attempt to take control. Origin, intent and the circumstances of use all matter.
A sentence cannot reliably supply that context by its wording alone. An assistant needs to understand the role in which the content is being returned.
The hypothesis we tested
The Lab evaluated several classifiers on synthetic data, using thresholds established before measurement. The development corpus contains 180 documents, including malicious texts, difficult legitimate cases, quotations and intended instructions.
The target combines two requirements: detect enough dangerous cases while keeping ordinary memory usable. A classifier that marks everything as suspicious can achieve perfect recall. That does not make it an acceptable protection.
The candidates did not all cover exactly the same documents, and some processed different text windows. A useful comparison preserves those differences instead of turning their latency figures into a simple speed ranking.
The experiment also separates development from independent confirmation. A result used to understand a failure is not automatically sufficient to qualify a safeguard.
Why none of the candidates was sufficient: scope and limits
On the injection cases used for the main recall criterion, Qwen3-4B reached 0.27 and Granite Guardian 0.24. The multilingual BERT control marked every document as suspicious at the examined threshold. The first candidates missed too much; the control stopped distinguishing legitimate use from risk.
The reference model, Qwen3.8-Flash, detected the injections in this development corpus. It also marked 12 of the 42 difficult legitimate texts as suspicious and confused intended instructions with attacks.
The lesson extends beyond model selection. An overly broad target class can reward behaviour that blocks exactly what users want to preserve. A more capable model does not automatically repair the question it is being asked.
These observations come from a development corpus without an independent confirmation test qualifying a protection. They do not establish detection rates for arbitrary documents or real organisational workloads.
Provenance matters alongside wording
Other Lab experiments found that rewriting can remove the visible form of a malicious instruction and present it as an established fact. Looking only for suspicious wording in the final text then leaves its history out of the assessment.
Provenance helps distinguish a person’s statement from imported material, a quotation or an AI reformulation. It does not establish that a claim is true. It makes its origin and the authority it should receive easier to examine.
The distinction between data and instructions must also be clear to the receiving assistant. No isolated detector replaces access rights, source controls, version history or the ability to ask for human confirmation when authority is uncertain.
These mechanisms address different questions. Knowing that a user may read a document does not mean every sentence in it is authorised to direct the assistant.
What this clarifies for Nawame
Nawame is portable, sovereign memory. Its role is to preserve context, reasons and provenance, then return them to authorised assistants and software. That role does not require its own language models to systematically interpret clients’ private content.
Running a model locally does not remove the issue. It remains necessary to establish what processing is authorised, who controls the data, what derived representations are produced and how corrections propagate.
Our experiments encourage us to define the guarantees before choosing a classifier. Sovereignty depends on trust boundaries and verifiable controls. This connects directly with our work on targeted passages: making a piece of text visible never gives it automatic authority as an instruction.
Related reading: How AI memory can become a prompt injection channel and The difference between a memory, a fact, a decision, a hypothesis and an instruction.


