Policy Compiler for Secure Agentic Systems

LLM-based agents are increasingly being deployed in contexts requiring complex authorization policies: customer service protocols, approval workflows, data access restrictions, and regulatory compliance. Embedding these policies in prompts provides no enforcement guarantees. We present PCAS, a Policy Compiler for Agentic Systems that provides deterministic policy enforcement. Enforcing such policies requires tracking information flow across agents, which linear message histories cannot capture. Instead, PCAS models the agentic system state as a dependency graph capturing causal relationships among events such as tool calls, tool results, and messages. Policies are expressed in a Datalog-derived language, as declarative rules that account for transitive information flow and cross-agent provenance. A reference monitor intercepts all actions and blocks violations before execution, providing deterministic enforcement independent of model reasoning. PCAS takes an existing agent implementation and a policy specification, and compiles them into an instrumented system that is policy-compliant by construction, with no security-specific restructuring required. We evaluate PCAS on three case studies: information flow policies for prompt injection defense, approval workflows in a multi-agent pharmacovigilance system, and organizational policies for customer service. On customer service tasks, PCAS improves policy compliance from 48% to 93% across frontier models, with zero policy violations in instrumented runs.

Through the Stealth Lens: Rethinking Attacks and Defenses in RAG

Retrieval-augmented generation (RAG) systems are vulnerable to attacks that inject poisoned passages into the retrieved set, even at low corruption rates. We show that existing attacks are not designed to be stealthy, allowing reliable detection and mitigation. We formalize stealth using a distinguishability-based security game. If a few poisoned passages are designed to control the response, they must differentiate themselves from benign ones, inherently compromising stealth. This motivates the need for attackers to rigorously analyze intermediate signals involved in generation$\unicode{x2014}$such as attention patterns or next-token probability distributions$\unicode{x2014}$to avoid easily detectable traces of manipulation. Leveraging attention patterns, we propose a passage-level score$\unicode{x2014}$the Normalized Passage Attention Score$\unicode{x2014}$used by our Attention-Variance Filter algorithm to identify and filter potentially poisoned passages. This method mitigates existing attacks, improving accuracy by up to $\sim 20 \%$ over baseline defenses. To probe the limits of attention-based defenses, we craft stealthier adaptive attacks that obscure such traces, achieving up to $35 \%$ attack success rate, and highlight the challenges in improving stealth.

Attacking Byzantine Robust Aggregation in High Dimensions

Training modern neural networks or models typically requires averaging over a sample of high-dimensional vectors. Poisoning attacks can skew or bias the average vectors used to train the model, forcing the model to learn specific patterns or avoid learning anything useful. Byzantine robust aggregation is a principled algorithmic defense against such biasing. Robust aggregators can bound the maximum bias in computing centrality statistics, such as mean, even when some fraction of inputs are arbitrarily corrupted. Designing such aggregators is challenging when dealing with high dimensions. However, the first polynomial-time algorithms with strong theoretical bounds on the bias have recently been proposed. Their bounds are independent of the number of dimensions, promising a conceptual limit on the power of poisoning attacks in their ongoing arms race against defenses. In this paper, we show a new attack called HIDRA on practical realization of strong defenses which subverts their claim of dimension-independent bias. HIDRA highlights a novel computational bottleneck that has not been a concern of prior information-theoretic analysis. Our experimental evaluation shows that our attacks almost completely destroy the model performance, whereas existing attacks with the same goal fail to have much effect. Our findings leave the arms race between poisoning attacks and provable defenses wide open.

RETEXO: Scalable Neural Network Training over Distributed Graphs

Graph neural networks offer a promising approach to supervised learning over graph data. Graph data, especially when it is privacy-sensitive or too large to train on centrally, is often stored partitioned across disparate processing units (clients) which want to minimize the communication costs during collaborative training. The fully-distributed setup takes such partitioning to its extreme, wherein features of only a single node and its adjacent edges are kept locally with one client processor. Existing GNNs are not architected for training in such setups and incur prohibitive costs therein. We propose RETEXO, a novel transformation of existing GNNs that improves the communication efficiency during training in the fully-distributed setup. We experimentally confirm that RETEXO offers up to 6 orders of magnitude better communication efficiency even when training shallow GNNs, with a minimal trade-off in accuracy for supervised node classification tasks.