LLM Security Research · Lateos
Adversarial robustness evaluation using the IPI Taxonomy. 250 test cases across indirect prompt injection classes. Structural disclosure with defensive mitigations and validation patterns.
Overview
This assessment evaluates GPT-OSS-120B's resistance to indirect prompt injection (IPI) attacks — the class of attack where untrusted content in a model's context attempts to redirect or override its intended behavior. Testing covers 250 cases across 25 IPI classes with 3 delivery variants each (direct, obfuscated, embedded).
Results show 8% overall susceptibility, with critical findings concentrated in specific IPI classes (IPI-010, IPI-019, IPI-021, IPI-022). The model demonstrates strong resistance to most injection patterns but exhibits measurable vulnerabilities in certain architectural scenarios.