Berk Sunar

The rapid deployment of Large Language Models (LLMs) has created an urgent need for enhanced security and privacy measures in Machine Learning (ML). LLMs are increasingly being used to process untrusted text inputs and even generate executable code, often while having access to sensitive system controls. To address these security concerns, several companies have introduced guard models, which are smaller, specialized models designed to protect text generation models from adversarial or malicious inputs. In this work, we advance the study of adversarial inputs by introducing Super Suffixes, suffixes capable of overriding multiple alignment objectives across various models with different tokenization schemes. We demonstrate their effectiveness, along with our joint optimization technique, by successfully bypassing the protection mechanisms of Llama Prompt Guard 2 on five different text generation models for malicious text and code generation. To the best of our knowledge, this is the first work to reveal that Llama Prompt Guard 2 can be compromised through joint optimization. Additionally, by analyzing the changing similarity of a model's internal state to specific concept directions during token sequence processing, we propose an effective and lightweight method to detect Super Suffix attacks. We show that the cosine similarity between the residual stream and certain concept directions serves as a distinctive fingerprint of model intent. Our proposed countermeasure, DeltaGuard, significantly improves the detection of malicious prompts generated through Super Suffixes. It increases the non-benign classification rate to nearly 100%, making DeltaGuard a valuable addition to the guard model stack and enhancing robustness against adversarial prompt attacks.

We propose using reinforcement learning to address the challenges of discovering microarchitectural vulnerabilities, such as Spectre and Meltdown, which exploit subtle interactions in modern processors. Traditional methods like random fuzzing fail to efficiently explore the vast instruction space and often miss vulnerabilities that manifest under specific conditions. To overcome this, we introduce an intelligent, feedback-driven approach using RL. Our RL agents interact with the processor, learning from real-time feedback to prioritize instruction sequences more likely to reveal vulnerabilities, significantly improving the efficiency of the discovery process. We also demonstrate that RL systems adapt effectively to various microarchitectures, providing a scalable solution across processor generations. By automating the exploration process, we reduce the need for human intervention, enabling continuous learning that uncovers hidden vulnerabilities. Additionally, our approach detects subtle signals, such as timing anomalies or unusual cache behavior, that may indicate microarchitectural weaknesses. This proposal advances hardware security testing by introducing a more efficient, adaptive, and systematic framework for protecting modern processors. When unleashed on Intel Skylake-X and Raptor Lake microarchitectures, our RL agent was indeed able to generate instruction sequences that cause significant observable byte leakages through transient execution without generating any $μ$code assists, faults or interrupts. The newly identified leaky sequences stem from a variety of Intel instructions, e.g. including SERIALIZE, VERR/VERW, CLMUL, MMX-x87 transitions, LSL+RDSCP and LAR. These initial results give credence to the proposed approach.

In the past decade, many vulnerabilities were discovered in microarchitectures which yielded attack vectors and motivated the study of countermeasures. Further, architectural and physical imperfections in DRAMs led to the discovery of Rowhammer attacks which give an adversary power to introduce bit flips in a victim's memory space. Numerous studies analyzed Rowhammer and proposed techniques to prevent it altogether or to mitigate its effects. In this work, we push the boundary and show how Rowhammer can be further exploited to inject faults into stack variables and even register values in a victim's process. We achieve this by targeting the register value that is stored in the process's stack, which subsequently is flushed out into the memory, where it becomes vulnerable to Rowhammer. When the faulty value is restored into the register, it will end up used in subsequent iterations. The register value can be stored in the stack via latent function calls in the source or by actively triggering signal handlers. We demonstrate the power of the findings by applying the techniques to bypass SUDO and SSH authentication. We further outline how MySQL and other cryptographic libraries can be targeted with the new attack vector. There are a number of challenges this work overcomes with extensive experimentation before coming together to yield an end-to-end attack on an OpenSSL digital signature: achieving co-location with stack and register variables, with synchronization provided via a blocking window. We show that stack and registers are no longer safe from the Rowhammer attack.

In the past decade, many vulnerabilities were discovered in microarchitectures which yielded attack vectors and motivated the study of countermeasures. Further, architectural and physical imperfections in DRAMs led to the discovery of Rowhammer attacks which give an adversary power to introduce bit flips in a victim's memory space. Numerous studies analyzed Rowhammer and proposed techniques to prevent it altogether or to mitigate its effects. In this work, we push the boundary and show how Rowhammer can be further exploited to inject faults into stack variables and even register values in a victim's process. We achieve this by targeting the register value that is stored in the process's stack, which subsequently is flushed out into the memory, where it becomes vulnerable to Rowhammer. When the faulty value is restored into the register, it will end up used in subsequent iterations. The register value can be stored in the stack via latent function calls in the source or by actively triggering signal handlers. We demonstrate the power of the findings by applying the techniques to bypass SUDO and SSH authentication. We further outline how MySQL and other cryptographic libraries can be targeted with the new attack vector. There are a number of challenges this work overcomes with extensive experimentation before coming together to yield an end-to-end attack on an OpenSSL digital signature: achieving co-location with stack and register variables, with synchronization provided via a blocking window. We show that stack and registers are no longer safe from the Rowhammer attack.

Bitslicing is a software implementation technique that treats an N -bit processor datapath as N parallel single-bit datapaths. Bitslicing is particularly useful to implement data-parallel algorithms, algorithms that apply the same operation sequence to every element of a vector. Indeed, a bit-wise processor instruction applies the same logical operation to every single-bit slice. A second benefit of bitsliced execution is that the natural spatial redundancy of bitsliced software can support countermeasures against fault attacks. A k -redundant program on an N -bit processor then runs as N/k parallel redundant slices. In this contribution, we combine these two benefits of bitslicing to implement a fault countermeasure for the number-theoretic transform (NTT) . The NTT efficiently implements a polynomial multiplication. The internal symmetry of the NTT algorithm lends itself to a data-parallel implementation, and hence it is a good candidate for the redundantly bitsliced implementation. We implement a redundantly bitsliced NTT on an advanced 667MHz ARM Cortex-A9 processor, and study the fault coverage for the protected NTT under optimized electromagnetic fault injection (EMFI) . Our work brings two major contributions. First, we show for the first time how to develop a redundantly bitsliced version of the NTT. We integrate the protected NTT into a full Dilithium signature sequence. Second, we demonstrate an EMFI analysis on a prototype implementation of the Dilithium signature sequence on ARM Cortex-M9. We perform a detailed EM fault-injection parameter search to optimize the location, intensity and timing of injected EM pulses. We demonstrate that, under optimized fault injection parameters, about 10% of the injected faults become potentially exploitable. However, the redundantly bitsliced NTT design is able to catch the majority of these potentially exploitable faults, even when the remainder of the Dilithium algorithm as well as the control flow is left unprotected. To our knowledge, this is the first demonstration of a bitslice-redundant design of the NTT that offers distributed fault detection throughout the execution of the algorithm.

Security critical software, e.g., OpenSSL, comes with numerous side-channel leakages left unpatched due to a lack of resources or experts. The situation will only worsen as the pace of code development accelerates, with developers relying on Large Language Models (LLMs) to automatically generate code. In this work, we explore the use of LLMs in generating patches for vulnerable code with microarchitectural side-channel leakages. For this, we investigate the generative abilities of powerful LLMs by carefully crafting prompts following a zero-shot learning approach. All generated code is dynamically analyzed by leakage detection tools, which are capable of pinpointing information leakage at the instruction level leaked either from secret dependent accesses or branches or vulnerable Spectre gadgets, respectively. Carefully crafted prompts are used to generate candidate replacements for vulnerable code, which are then analyzed for correctness and for leakage resilience. From a cost/performance perspective, the GPT4-based configuration costs in API calls a mere few cents per vulnerability fixed. Our results show that LLM-based patching is far more cost-effective and thus provides a scalable solution. Finally, the framework we propose will improve in time, especially as vulnerability detection tools and LLMs mature.