Archive for the ‘Blog’ Category

Other attempts to identify vulnerabilities in GPT-OSS In the process of working on the gpt-oss-20b red team, CCSW attempted to find vulnerabilities in a few areas. Unfortunately, these areas of research did not end up resulting in findings which could be reported for the competition. Despite these shortcomings, we feel that they are worth talking… » Read More

gpt-oss-20b is not always faithful in its chain-of-thought In the previous blog post, we showcased a Role in Prompt (RiP) attack where malicious user input and tool output can cause prompt injection and bypass the alignment safeguards in OpenAI’s newest gpt-oss-20b model. This finding was uncovered as part of the Caesar Creek Software’s team research… » Read More

We participated in a Kaggle competition to red team the OpenAI gpt-oss-20b model. The following series is a detailed look at our thoughts that encompassed our submission. Work conducted by Danny L., Huy Chi Dai, Cole L., John H., and Zack B. Bypassing Instruction Hierarchy and Policies in gpt-oss-20b  Introduction At the start of August… » Read More

Previously, we discussed pre-processing captured data for restoration. In this post, we discuss the audio restoration process using Generative Adversarial Networks (GANs) to aid in recovery. Audio Restoration Introduction In the previous post, we discussed the methods we used to pre-process the accelerometer and gyroscope data. In this blog post, we will discuss the machine… » Read More

After initial testing, we started to expand with our new tests and implementations. Preprocessing Introduction Previously, we explored the feasibility of the sensor attack vector, and found that it would be best realized through a native Android app. In this post, we will pre-process the data recovered from our application and discuss the decisions we… » Read More