Project Everest's verified TLS stack in F* took ~100 person-years and 1.5M lines of code/proof. Since then, we enlisted agents to build verified TLS 1.3 and crypto implementations in 6 weeks each, and X.509 in 6 hours. In line with similar advances in Lean and Rocq, such experiments highlight three challenges for agentic proof-oriented programming.

1/ How to audit specifications? We leveraged agents to implement and prove 52 CLRS algorithms and data structures in Pulse, the separation logic framework for F*; they initially came with imprecise specs. We introduce Small Proof-Oriented Tests (SPOTs) to expose weaknesses through proven clients. SPOTs, abstraction and ghost state support auditing and strengthening specs, including complexity analysis.

2/ How to prove properties beyond memory safety and functional correctness? We leveraged agents to produce Atlas, a Pulse implementation of TLS 1.3, X.509 and NIST crypto, verified for key agreement. ~500K Pulse lines yield 100K C lines; ~500 theorem lines, rubrics and state-machine specs streamline review. Proven optimizations close the gap with OpenSSL.

3/ How to verify existing code at scale? We introduce PAL, a Proof Annotation Language for C, along with a compiler translating annotated C into Pulse with separation logic specs and proofs. Agents annotate C in place, enabling verification of Linux's loopback driver and Microsoft's MSQUIC.

In regression and classification tasks with tabular datasets, most of the literature on explanation in ML (e.g. SHAP, permutation importance) has focused on predictive attribution: does feature X_i help predict outcome Y? At its heart, predictive attribution captures correlation, not causation. We instead ask a causal question: would manipulating feature X_i change outcome Y? We define a natural quantity to answer this question, the feature intervention effect: Δ_i = E[Y | do(X_i=+1)] − E[Y | do(X_i=−1)]. In general, estimating this quantity from observational data alone would entail strong causal assumptions. We propose FADO (Feature-wise Amortized Do-Operators), an amortized inference method using prior-fitted networks trained on synthetic tabular datasets generated by structural causal models. During training, the synthetic data generating process is known, so we can compute ground truth intervention effects Δ_i, and train a model to predict them at test time from a purely observational tabular dataset. As we demonstrate through extensive experiments, our new causal explainability method is easy to train and provides better explanations under a range of settings for which standard predictive attribution fails, including proxy variables, confounders, colliders, and placebos. We evaluate it on real, synthetic and semi-synthetic datasets, with and without hidden confounders, and compare it to existing predictive attribution methods and state-of-the-art causal inference methods.

As AI agents increasingly engage in high-stakes decision making, it is essential to assess their accountability in ways that are both fair and interpretable. This involves explaining expected or realized outcomes of multi-agent systems and attributing responsibility for those outcomes to the participating agents. Addressing these challenges is key to fostering societal trust and easing the adoption of AI decision makers. This thesis investigates accountability in multi-agent sequential decision making. We develop methods to attribute responsibility for observed outcomes and overall system performance, design efficient approximation algorithms for otherwise intractable attribution problems, and introduce causal tools to explain how agents’ decisions influence outcomes. Together, these contributions establish theoretical foundations and practical tools for accountable decision making, drawing on and integrating insights from causality, multi-agent reinforcement learning and game theory.

We build secure software on a seemingly obvious assumption: the processor executes instructions correctly and enforces the boundaries between programs, operating systems, and virtual machines. But what if the CPU itself breaks these rules? Over the past decade, research has found increasingly powerful ways in which processors violate the abstractions on which software relies. What started with tiny timing differences revealing secrets led to attacks such as Spectre and Meltdown, and more recently to CPU bugs that can directly corrupt data, bypass isolation, or expose information across security boundaries.

In this talk, I will show how seemingly harmless implementation details and useful optimizations turn into practical security vulnerabilities. Using examples from our research across Intel, AMD, and RISC-V processors, I will talk about how we discover and understand CPU vulnerabilities and ultimately turn them into security problems. The broader lesson goes beyond processor security: abstractions are only as strong as the layers implementing them. If we want trustworthy computer systems, hardware must increasingly be tested, analyzed, and secured with the same rigor that we already apply to software.

Anomaly detection is one of the fundamental topics in machine learning and artificial intelligence. The aim is to identify instances that deviate from the norm, so-called anomalies. Such anomalies arise in many scenarios, ranging from attacks on computer or energy networks to critical faults in industrial systems or rare tumors in medical imaging data. In this talk, I will first provide an overview of the field of anomaly detection, with an emphasis on approaches based on deep neural networks. I will then present recent algorithms and theoretical advances in deep anomaly detection. In particular, I will discuss why certain methods work and how their behavior can be better understood using principled analysis. Furthermore, I will highlight recent developments in anomaly detection for time series data, including new benchmark datasets and theoretically-grounded evaluation metrics. Finally, I will conclude with an outlook on promising future research directions in anomaly detection.

Reinforcement learning (RL) has emerged as a fundamental approach to decision-making in machine learning, with applications across a wide range of real-world domains, and several practical extensions, such as multi-agent RL (MARL) and RL from human feedback (RLHF). Despite the growing successful applications of these systems, there exists an inherent threat when it comes to applying them in the real world, where ill-intentioned third parties may intervene in both their training process and deployment. This typically has catastrophic consequences, where even small and inexpensive perturbations to the environment may cause the system to substantially diverge from the desired behavior. The purpose of this thesis is to provide a thorough investigation of various adversarial attacks and robustness against such attacks to (MA)RL and (MA)RLHF systems. In particular, we study training-time and test-time attacks in MARL and propose algorithms that are shown to be provably robust against such attacks. Beyond MARL, we establish fundamental statistical results for RLHF, provide a rigorous characterization of poisoning attacks to RLHF, and develop robust extensions for MARLHF. Our work is centered around robust algorithmic approaches with provable guarantees under common assumptions.