Incubated ML Exploits: Backdooring ML Pipelines with Input Handling Bugs
Demonstrates incubated ML exploits using LangSec principles to inject backdoors into machine learning pipelines via model serialization bugs.
Summary
The ML Pipeline Attack Surface
Machine learning models are not isolated files; they are complex artifacts generated and processed by extensive toolchains, serialization libraries, and automated pipelines.
Language Theoretic Security (LangSec) in ML
Applying LangSec principles reveals that model parsers frequently contain input-handling and deserialization vulnerabilities that can be exploited during model ingestion.
Constructing Polyglot and Ambiguous Artifacts
Researchers crafted malicious polyglot and ambiguous model files that evade static scanners while executing arbitrary backdoor triggers when loaded by ML frameworks.
Advancing Fickling and ML Security Guidelines
Contributions to Fickling, a pickle security tool for ML, along with practical development guidelines, help practitioners secure model serialization workflows.
Related talks
Creator Stage·40 min
Chloé Messdaghi · Kasimir Schulz
Explores the ethical challenges of AI vulnerability disclosure, covering data bias exploitation and responsible reporting standards.
Creator Stage·22 min
D. Jurado · J. Nogue
Chronicles the deployment of XBOW, an autonomous AI offensive agent that discovered over 200 open-source vulnerabilities and submitted 1,000 bug bounty reports.
Creator Stage·26 min
Patrick Walsh · Bob Wall
Explains embedding inversion attacks against generative AI vector databases and introduces open-source vector encryption for private search.