Language Models have non-patchable vulnerabilities given shared lineage and function
(e.g. Transformers, Common Crawl, etc).
Attacks can be Automatically Customized by malicious actors to affect specific ends
(e.g. Privilege Escalation, Data Extraction, etc).
from the people that brought you ZeroDay.Tools; code via Github w/ executive summary available
Latent Space Tools help conceptualize, visualize, and subsequently operationalize the necessary architecture and software components for secure LLM Deployment & Monitoring.
Latent Space Tools are made available under the Apache 2 license via Github
1) Prompt Injection Detection & Mitigation
2) Service Denial & Performance Monitoring
3) Topic / Sentiment Modeling x Vector Comparisons & Cluster Definition
4) Attack Mitigation, Appending (Un)Certainty & Response Non-Conformity
5) Heatmaps x Dimensionality Drift via Conformal Prediction Intervals
Note: Actively developing models designed as additional pre-processing to differentiate attack strings vs parameterized URLs; also looking to develop membership and attribute inference attacks as pipelines to affect point-forward GDPR compliant 'forgetting' for DNNs utilizing open-source tools like WeightWatcher.ai for layer-specific validation.
based on A16Z's Reference Architecture; now with grounding
more details available on GitHub
Given that a latent space generally represents a reduced dimensionality compared to the feature space, we expect the 'aggregate' dimensions to move around more than their component parts.
That said, the chosen dimensions should represent meaningful metrics worth monitoring. Hence, the importance of conceptualizing, monitoring, and forecasting changes to those values.
Latent Space Tools extensively leverage the concept of conformal prediction; whereby previous outputs better predict future outputs than do Bayesian priors or assumptions.