Apple's DRM stack protects media content and applications. Across the industry, external attackers increasingly use LLMs — interactively or via automated/agentic pipelines — to accelerate reverse-engineering of software protections; this is a general and growing trend, not specific to any one target. This internship addresses that trend directly: developing, from first principles, inputs that measurably perturb an LLM's internal state (activations, attention distributions, representation geometry) in ways that degrade its ability to usefully assist an attacker analyzing protected code/binaries.
This internship addresses that trend directly: developing, from first principles, inputs that measurably perturb an LLM's internal state (activations, attention distributions, representation geometry) in ways that degrade its ability to usefully assist an attacker analyzing protected code/binaries.
Formulate the problem mathematically: define what "confusion" or degraded internal state means in terms of measurable quantities (activation/representation drift, attention entropy, logit distribution shift, loss landscape behavior) rather than qualitative output judgments.
Develop gradient-based and optimization-based methods for crafting adversarial inputs (e.g. embedding-space perturbation, discrete token optimization à la GCG/greedy coordinate search, universal adversarial triggers) targeting the internal-state objectives defined above.
Analyze transformer internals (attention matrices, residual stream, activation subspaces) to identify structural properties — e.g. spectral characteristics, representation collapse, direction vectors linked to specific behaviors — that can be exploited or reinforced.
Study whether techniques derived on open-weight models (full white-box access) transfer to closed-weight models via black-box/transfer-based optimization, and quantify the transfer gap mathematically.
Build a rigorous evaluation methodology: statistically sound metrics, ablations isolating which mathematical property of the input drives the effect, and robustness checks across model families and versions.
Assess durability: whether effects derived from current architectures are expected to hold under fine-tuning, distillation, or future model versions, based on the underlying mathematical mechanism rather than empirical trial-and-error.
Document the mathematical framework, methods, and results internally for Apple's use only; no external publication.
Currently completing a Master's degree or enrolled in a PhD, in Computer Science, Applied Mathematics, or a related field, with strong theoretical grounding.
Strong foundations in linear algebra, optimization (convex and non-convex), and probability/information theory, and ability to apply them to deep learning internals rather than just use libraries.
Solid understanding of transformer architecture internals (attention mechanics, residual stream, embeddings) at a mathematical level.
Hands-on experience with gradient-based adversarial attack techniques (FGSM/PGD-style methods, discrete optimization for token sequences, universal adversarial perturbations).
Proficiency in Python and PyTorch, including writing custom gradient computations/hooks into model internals (not just using high-level APIs).
Research experience in adversarial machine learning or mechanistic interpretability (activation patching, representation engineering, steering vectors).
Familiarity with reverse engineering / binary analysis, to reason about what a real attacker workflow needs from the LLM.
Coursework or research in spectral methods, information theory, or statistical learning theory applied to neural network analysis.
Prior publications, competition results (e.g. adversarial ML challenges), or open-source contributions demonstrating this kind of mathematical adversarial ML work.
Job description provided by Apple.
View the original posting