Omni-Scanner: Detecting Hallucinations and Manipulation in Large Language Models Using Topological Data Analysis

Omni-Scanner is an open-source forensic auditing tool that uses Topological Data Analysis (TDA) and the principle of semantic invariance to provide mathematically objective verification for hallucination detection, gaslighting identification, and structural manipulation analysis of large language models. It aims to address the urgent need for AI honesty detection by analyzing the internal representation structure of models to reveal real cognitive patterns, going beyond the surface observations of enterprise filtering layers.

Background: Urgent Need for AI Honesty Detection

With the widespread application of large language models (LLMs) across various industries, an increasingly critical question has emerged: How can we ensure these AI systems are telling the truth? Enterprise-level LLMs are usually equipped with complex filtering mechanisms and safety layers, which, while designed to prevent harmful outputs, may also mask the model's true behavioral patterns. More problematic is that modern LLMs exhibit a phenomenon known as "gaslighting": when faced with errors, they may confidently stick to their wrong positions, even making users doubt their own judgment.

Traditional model evaluation methods mainly rely on manual annotation and benchmark testing, but these methods have obvious limitations: they often only capture surface behaviors and struggle to identify deep structural manipulation; they depend on human judgment, and humans themselves may be misled by the AI's fluent expressions; more importantly, these methods usually can only observe the model outside the enterprise filtering layer, unable to access the model's internal true representations.

Technical Principle: The Power of Topological Data Analysis

Topological Data Analysis is a mathematical framework derived from algebraic topology, which excels at extracting robust shape features from high-dimensional data. In the context of Omni-Scanner, TDA is used to analyze the geometric structure of internal activation vectors of LLMs. Specifically, when a model processes input, each layer generates high-dimensional activation vectors. These vectors form specific "shapes" or manifolds in high-dimensional space. TDA uses techniques like Persistent Homology to capture the topological features of these shapes—such as the number of connected components, ring structures, and voids. The key point is that these topological features are robust to noise and minor perturbations, while also revealing the deep structure behind the data. When a model generates hallucinations, its internal activation patterns exhibit specific abnormal topological features. For example, the model may establish false high-dimensional connections between semantically unrelated concepts, or create unnatural "bridges" between concept regions that should remain separate. Omni-Scanner establishes a detection mechanism by learning the differences in topological features between "honest answers" and "hallucinatory answers.

Semantic Invariance: A Cross-Language Standard for Truth

Semantic invariance is another core pillar of Omni-Scanner. This principle is based on a profound insight: true knowledge and concepts should have stability across languages and forms of expression. If a concept is true in English, it should maintain its truth structure in Chinese, French, or any other language. In practice, Omni-Scanner uses the alignment properties of multilingual embedding spaces to detect semantic drift. When a model produces significantly different internal representations for different linguistic expressions of the same concept, this may imply that the model does not truly "understand" the concept but is performing surface pattern matching. More seriously, if the model gives contradictory answers to the same factual question in different languages, this constitutes clear evidence of gaslighting. By constructing a semantic invariance metric, Omni-Scanner can quantify the cross-language consistency of model answers, thereby identifying "local truths" that only hold in specific languages or forms of expression—these are often breeding grounds for hallucinations and manipulation.

Application Scenarios: From Auditing to Protection

Omni-Scanner is designed to provide value to multiple stakeholders. For AI security researchers, it offers a new tool for quantitative analysis of model behavior; for enterprise users, it can serve as an internal auditing mechanism to ensure deployed models meet honesty standards; for regulatory bodies, it provides a technology-neutral verification method that can be compared across different vendors and model architectures.

Specific applications include:

Hallucination Detection: Identifying the fabrication behavior of models when they lack real knowledge, especially critical in professional fields (medicine, law, engineering).

Gaslighting Identification: Detecting defensive, misleading, or denial behavior patterns exhibited by models when faced with corrections—this is particularly important in sensitive applications like mental health and education.

Structural Bias Auditing: Identifying implicit stereotypes and biases by analyzing the topological structure of model internal representations, which may not be visible in standard fairness metrics.

Adversarial Testing: Evaluating the vulnerability of models to malicious prompts, especially "jailbreak" attempts that try to bypass security mechanisms.

Technical Implementation and Open-Source Significance

Omni-Scanner is released as an open-source project, which in itself has significant value. In the current AI field, most model evaluation tools are either developed internally by large tech companies and not made public, or can only touch the surface behavior of models. Open-source allows the research community to independently verify, improve, and extend these detection methods, forming a balancing force against the interpretive power of a single vendor. The project is implemented in Python, integrating mature TDA libraries (such as GUDHI, Ripser) and deep learning frameworks. Its modular design allows users to customize detection processes for specific model architectures (Transformer, Mamba, etc.) and analysis goals.

Limitations and Future Directions

Although Omni-Scanner provides a powerful technical framework, users should be aware of its current limitations. Topological analysis has high computational costs and may require significant computing resources for ultra-large-scale models; in addition, this method mainly analyzes the model's internal activations, so its applicability to black-box API models is limited.

Future development directions include: developing more efficient approximation algorithms to reduce computational overhead; expanding support for multimodal models (vision-language models); establishing standardized benchmark datasets for hallucinations and gaslighting; and deep integration with model interpretability technologies.

Conclusion

Omni-Scanner represents an important attempt in the field of AI security: using mathematical objectivity to counter model deceptiveness. In a world increasingly dependent on AI-assisted decision-making, having tools to independently verify AI honesty is no longer a luxury but a necessity. Through Topological Data Analysis and the principle of semantic invariance, Omni-Scanner provides us with a window into the "inner world" of AI—where the truth is no longer hidden by fluent wording but presented as it is in its structural form.