We introduce PurpCode, the first post-training recipe for training safe code reasoning models towards generating secure code and defending against malicious cyberactivities. PurpCode trains a reasoning model in two stages: (i) Rule Learning, which explicitly teaches the model to reference cybersafety rules to generate vulnerability-free code and to avoid facilitating malicious cyberactivities; and (ii) Reinforcement Learning, which optimizes model safety and preserves model utility through diverse, multi-objective reward mechanisms. To empower the training pipelines with comprehensive cybersafety data, we conduct internal red-teaming to synthesize comprehensive and high-coverage prompts based on real-world tasks for inducing unsafe cyberactivities in the model. Based on PurpCode, we develop a reasoning-based coding model, namely PurpCode-32B, which demonstrates state-of-the-art cybersafety, outperforming various frontier models. Moreover, our alignment method decreases the model overrefusal rates in both general and cybersafety-specific scenarios, while preserving model utility in both code generation and common security knowledge.

Static binary analysis is critical to various security tasks such as vulnerability discovery and malware detection. In recent years, binary analysis has faced new challenges as vendors of the Internet of Things (IoT) and Industrial Control Systems (ICS) continue to introduce customized or non-standard binary formats that existing tools cannot readily process. Reverse-engineering each of the new formats is costly as it requires extensive expertise and analysts' time. In this paper, we investigate the first step to automate the analysis of non-standard binaries, which is to recognize the bytes representing “code” from “data” (i.e., data-code separation). We propose Loadstar, and its key idea is to use the abundant labeled data from standard binaries to train a classifier and adapt it for processing unlabeled non-standard binaries. We use a pseudo-label-based method for domain adaption and leverage knowledge-inspired rules for pseudo-label correction, which serves as the guardrail for the adaption process. A key advantage of the system is that it does not require labeling any non-standard binaries. Using three datasets of non-standard PLC binaries, we evaluate Loadstar and show it outperforms existing tools in terms of both accuracy and processing speed. We will share the tool (open source) with the community.

Patient outreach enables timely communication between patients and healthcare providers but is vulnerable to phishing/spoofing attacks. In this paper, we work with a U.S.-based healthcare provider to design an inclusive method to address this threat. We present VeriSMS which allows patients to call a voice agent to verify whether the received (sensitive) messages are indeed sent by their healthcare provider. We design the system to be inclusive: it is accessible to patients who only have access to SMS and phone call capabilities. We perform a two-part user study to refine the system design (N=15) and confirm users can correctly understand the system and use it to identify spoofed/phishing messages (N=35). A key insight from our study is to not exclusively optimize for strong security but to tailor the designs based on user habits. Our result confirms the effectiveness and usability of VeriSMS and its ability to significantly increase adversaries’ costs.

Significant efforts have been investigated to develop machine learning (ML) based tools to support security operations. However, they still face key challenges in practice. A generally perceived weakness of machine learning is the lack of explanation, which motivates researchers to develop machine learning explanation techniques. However, it is not yet well understood how security practitioners perceive the benefits and pain points of machine learning and corresponding explanation methods in the context of security operations. To fill this gap and understand "what is needed", we conducted semi-structured interviews with 18 security practitioners with diverse roles, duties, and expertise. We find practitioners generally believe that ML tools should be used in conjunction with (instead of replacing) traditional rule-based methods. While ML’s output is perceived as difficult to reason, surprisingly, rule-based methods are not strictly easier to interpret. We also find that only few practitioners considered security (robustness to adversarial attacks) as a key factor for the choice of tools. Regarding ML explanations, while recognizing their values in model verification and understanding security events, practitioners also identify gaps between existing explanation methods and the needs of their downstream tasks. We collect and synthesize the suggestions from practitioners regarding explanation scheme designs, and discuss how future work can help to address these needs.

Malicious manipulations on Industrial Control Systems (ICSs) endanger critical infrastructures, causing unprecedented losses. State-of-the-art research in the discovery and exploitation of vulnerability typically assumes full visibility and control of the industrial process, which in real-world scenarios is unrealistic. In this work, we investigate the possibility of an automated end-to-end attack for an unknown control process in the constrained scenario of infecting just one industrial computer. We create databases of human-machine interface images, and Programmable Logic Controller (PLC) binaries using publicly available resources to train machine-learning models for modular and granular fingerprinting of the ICS sectors and the processes, respectively. We then explore control-theoretic attacks on the process leveraging common/ubiquitous control algorithm modules like Proportional Integral Derivative blocks using a PLC binary reverse-engineering tool, causing stable or oscillatory deviations within the operational limits of the plant. We package the automated attack and evaluate it against a benchmark chemical process, demonstrating the feasibility of advanced attacks even in constrained scenarios.

Attacks on Industrial Control Systems (ICS) are increasingly targeting field devices and the firmware that instruments their operation. Securing the firmware images and their update procedure has, therefore, become an important challenge. This is especially true for widely deployed legacy devices which are not equipped with the necessary security mechanisms/capabilities. In this paper, we address the problem by reverse engineering PLC firmware update tools to build a device that ensures the integrity and authenticity of firmware updates, before allowing them to be flashed onto a field device. Our tool is directly connected to field devices and consists of a firmware signing mechanism, a PLC emulation module, and a payload detection classifier – all integrated in a bump-in-the-wire device, SNIFU. SNIFU monitors serial traffic sent to the PLC for firmware update commands. When it identifies such commands, it emulates a PLC, capturing the entire firmware image and verifying it before relaying it to the PLC. We implement and evaluate a prototype of SNIFU using a Raspberry Pi, that secures the update process of a commercial PLC by Wago.