Cyber Risks · Trail of Bits Blog

Our response to the US Army’s RFI on developing AIBOM tools

Summary: The US Army's PEO IEW&S issued an RFI on automating production of an AI bill of materials as part of Project Linchpin, which aims to build an operational pipeline for deploying AI/ML to intelligence, cyber, and electronic warfare systems. Trail of Bits' response argues the proposed AIBOM concept cannot deliver a complete security audit because aspects of model training and use cannot be captured statically, omits AI/ML-specific hardware via an HBOM, and misses dependencies such as ML-as-a-service prediction APIs.

Bottom line: The Army's own concept would not capture whether an attacker controlled the order in which a model ingests training data - a data poisoning vector - and the vulnerability databases it would rely on still lack a strong definition of what an AI/ML vulnerability is.

Read the original at Trail of Bits Blog ↗

Topics
AI Security
Related