Security papers — 2026-10-21

The Morning Preprint — this day as one edition: the day's papers, the five things worth knowing, and the episodes behind them.

Today’s work focuses on understanding the true financial cost of corporate data breaches because knowing the social cost helps frame the severity beyond just legal settlements. This is explored by looking at how much it actually costs victims, including lost time and healthcare expenses, which AKMPC tries to estimate by examining direct financial losses and increased identity theft incidents after a major event. This helps show that while large settlements exist, the marginal damage of individual records seems to lessen over time.

Another important area involves how to detect threats in complex systems, specifically agentic AI. Foundations for Investigations of Agentic AI Systems from the Forensic Analysis of OpenClaw explores reconstructing the internal state and actions of personal AI assistants after they interact with users. This is crucial because these systems are becoming common, and methods are needed to investigate them systematically.

There is also work on improving security operations by making alerts more useful. STRAL: What Adaptation Gains, Loses, and Costs in SOC Alert Screening investigates how adaptive screening controllers can reduce false alarms while trying to preserve important detections. This directly impacts the day-to-day work of security teams by balancing alert volume against missed threats.

Finally, there is research into digital authenticity and security measures where Authenticated Contradictions from Desynchronized Provenance and Watermarking formalizes a conflict that arises when cryptographic proofs of authorship clash with AI watermarks. This work shows how to close that gap through cross-layer auditing.

The most pressing issue right now is how to safely use machine learning in vehicles because this technology promises much for autonomous driving and connected services, but it brings huge privacy risks due to the massive data exchange involved. A comprehensive review of privacy-preserving machine learning in Internet of Vehicle applications provides the necessary foundation for understanding what has been tried and where we still need to go. This study first looks at existing surveys of various privacy-preserving machine learning techniques and how they are being integrated into IoV across different application scopes.

Building on that groundwork, the research then categorizes IoV applications into three main domains and analyzes the specific privacy challenges inherent in using machine learning within those areas. After establishing these fundamentals, the authors examine recent advancements in integrating various privacy-preserving techniques into IoV applications, discussing their frameworks and performance concerning privacy, utility, and efficiency. This leads into more specific technical work like Sequential Behavioral Watermarking for LLM Agents, which tries to embed signals directly into agent behavior rather than just generated text to establish provenance.

This behavioral watermarking approach is significant because it addresses the difficulty of tracking which agent produced a sequence of decisions when observing only the resulting behavior. However, the paper notes that treating each action step as an independent trial overlooks important trajectory structure, making these methods fragile when trajectories are altered or observed incompletely. This contrasts with other security concerns in AI systems, such as those discussed in Can You Trust the Vectors in Your Vector Database?, which shows that even vector databases can be vulnerable to black-hole attacks where malicious vectors near the geometric center of the embedding space attract most queries.

Moving toward broader system design, there is work on Redefining Hybrid Blockchains: A Balanced Architecture, which proposes a novel hybrid blockchain architecture using strategies like resource and node isolation to balance scalability and decentralization. This architecture simulates over one thousand transactions per second, suggesting a path toward more viable enterprise adoption for blockchain technology. Furthermore, in the realm of digital currency security, Perspectives on Objectives and Design Principles in Offline Payments with Central Bank Digital Currency discusses fundamental design principles for CBDCs focusing on access control and privacy by design.

Finally, there is research into securing data storage itself with Revisiting Shared Data Protection Against Key Exposure, which introduces a new computationally secure encryption-then-sharing scheme that protects data even when the encryption key is exposed. This work reduces the computational overhead of data encryption by half compared to current state-of-the-art methods.

The most significant takeaway from this research is how block withholding attacks behave under a pay-per-share scheme, which fundamentally changes the incentive landscape compared to older models. This work shows that classical block withholding attacks are not incentive compatible when miners are paid based on shares submitted rather than blocks found.

The core finding involves the all-out attack strategy where an adversary dedicates all its hashpower to a victim pool, submitting only partial proof-of-work shares while withholding full proof-of-work blocks. This strategy is particularly effective after the first difficulty adjustment because it allows the attacker to either generate more partial shares or earn a higher reward per share if the difficulty remains fixed. This results in a post-adjustment reward rate of alpha one minus alpha per target epoch for the attacker, which matches what is theoretically possible under optimal selfish mining with perfect network influence.

This result is important because it demonstrates that block withholding is substantially more profitable under pay-per-share than in traditional block-dependent payout schemes, even when compared to more advanced withholding variants. Honest miners actually earn the same rate per unit of hashpower as the attacker does, while the victim pool operator absorbs all costs by paying out for partial shares without receiving any full blocks in return. This dynamic is further contextualized by the analysis of All-out Attack: Optimal Block Withholding Under Pay-Per-Share Scheme, which contrasts these findings with earlier conclusions drawn under fixed difficulty in prior literature.

Today's papers

The papers

Important terms

Social Cost of Data Breaches
This concept frames data breach severity beyond legal fines by estimating real-world costs like lost time and healthcare expenses for victims, showing marginal damage decreases over time.
Agentic AI Forensics
Methods to reconstruct the internal state and actions of personal AI assistants after user interaction are crucial for systematically investigating these complex, common systems.
SOC Alert Screening (STRAL)
This research explores adaptive screening controllers that aim to reduce false alarms in security operations while preserving important threat detections, balancing alert volume against missed threats.
Authenticated Contradictions
This involves formalizing the conflict between cryptographic proofs of authorship and AI watermarks, requiring cross-layer auditing to close the gap between these security measures.