Researchers poison their own data when stolen by an AI to ruin results

Date:
Wed, 07 Jan 2026 17:20:00 +0000

Description:
Poisoned knowledge graphs can make the LLM hallucinate, rendering it useless
to the thieves.

FULL STORY

Researchers from universities in China and Singapore came up with a creative way to prevent the theft of data used in Generative AI .

Among other things, there are two important elements in todays Large Language Models (LLM): training data, and retrieval-augmented generation (RAG).

Training data teaches an LLM how language works and gives it broad knowledge
up to a cutoff point. It doesnt give the model access to new information, private documents, or fast-changing facts. Once training is done, that knowledge is frozen.

Replacing outdated gear

RAG, on the other hand, exists because many real questions depend on current, specific, or proprietary data (such as company policies, recent news,
internal reports, or niche technical documents). Instead of retraining the model every time data changes, RAG lets the model fetch relevant information
on demand and then write an answer based on it.

In 2024, Microsoft came up with GraphRAG - a version of RAG that organizes retrieved information as a knowledge graph instead of a flat list of
documents. This helps the model understand how entities, facts, and relationships connect to each other. As a result, the AI can answer more complex questions, follow links between concepts, and reduce contradictions
by reasoning over structured relationships rather than isolated text.

Since these knowledge graphs can be rather expensive, they could be targeted
by cybercriminals, nation-states, and other malicious entities.

In their research paper, titled Making Theft Useless: Adulteration-Based Protection of Proprietary Knowledge Graphs in GraphRAG Systems, authors
Weijie Wang, Peizhuo Lv, et al. proposed a defense mechanism called Active Utility Reduction via Adulteration, or AURA - which poisons the KGs, making
the LLM give wrong answers and hallucinate.

The only way to get correct answers is to have a secret key. The researchers said the system is not without its flaws, but that it works great in most
cases (94%).

"By degrading the stolen KG's utility, AURA offers a practical solution for protecting intellectual property in GraphRAG," the authors stated.

Via The Register

======================================================================
Link to news story: https://www.techradar.com/pro/security/researchers-poison-their-own-data-when- stolen-by-an-ai-to-ruin-results

$$
--- SBBSecho 3.28-Linux
* Origin: Capitol City Online (42:17/1)