Xavier Fernandes

Hojjat Aghakhani, Wei Dai, Andre Manoel et al. · microsoft-research, mit

With tools like GitHub Copilot, automatic code suggestion is no longer a dream in software engineering. These tools, based on large language models, are typically trained on massive corpora of code mined from unvetted public sources. As a result, these models are susceptible to data poisoning attacks where an adversary manipulates the model's training by injecting malicious data. Poisoning attacks could be designed to influence the model's suggestions at run time for chosen contexts, such as inducing the model into suggesting insecure code payloads. To achieve this, prior attacks explicitly inject the insecure code payload into the training data, making the poison data detectable by static analysis tools that can remove such malicious data from the training set. In this work, we demonstrate two novel attacks, COVERT and TROJANPUZZLE, that can bypass static analysis by planting malicious poison data in out-of-context regions such as docstrings. Our most novel attack, TROJANPUZZLE, goes one step further in generating less suspicious poison data by never explicitly including certain (suspicious) parts of the payload in the poison data, while still inducing a model that suggests the entire payload when completing code (i.e., outside docstrings). This makes TROJANPUZZLE robust against signature-based dataset-cleansing methods that can filter out suspicious sequences from the training data. Our evaluation against models of two sizes demonstrates that both COVERT and TROJANPUZZLE have significant implications for practitioners when selecting code used to train or tune code-suggestion models.

Kyle O'Brien, David Majercak, Xavier Fernandes et al.

Responsible deployment of language models requires mechanisms for refusing unsafe prompts while preserving model performance. While most approaches modify model weights through additional training, we explore an alternative: steering model activations at inference time via amplifying sparse autoencoder (SAE) features that mediate refusal. This work uncovers a fundamental tension between SAE steering-based safety improvements and general model capabilities. While feature steering successfully improves robustness against both single-turn and challenging multi-turn jailbreak attempts, we discover that this comes at a previously underexplored cost -- systematic degradation of performance across multiple benchmark tasks, even on safe inputs with no apparent connection to refusal behavior. This suggests that features mediating refusal may be more deeply entangled with general language model capabilities than previously understood. Our findings reveal important open questions about the nature of safety-relevant features in language models and the feasibility of isolating them for targeted intervention. While SAE-based steering shows promise as a flexible approach to enhancing language model safety, our results highlight the critical need to understand and address the mechanisms behind these capability tradeoffs before such techniques can be practically deployed.