RESEARCH
AI agents hijacked, memory sandboxes broken and side channels on implementations
OT and industrial securitySupply chain
A large conference's proceedings inflate the counts, so the jumps should be read with caution. The content is strong on AI agents (hijacking Android GUI agents, defenses for LLM ones), on defeating memory protections (V8 sandbox, kernel MTE) and on side channels in cryptographic implementations and TEEs.
This week's trends
AI & LLM Security · Steady
99 papers this week · average of the previous 4 weeks: 89.5
The count is near the average (99 versus 89.5), but the content clusters on agents and LLMs. One attack hijacks Android GUI agents by exploiting the gap between observation and action; other work proposes defenses against prompt injection, harmful reinforcement-learning fine-tuning and poisoning of code-generation models. For insurers, AI agents on corporate devices open vectors that usual controls miss.
Papers cited:
- Mind the Gap: Action Rebinding Attacks against Android GUI Agents
- LocalAlign: Enabling Generalizable Prompt Injection Defense via Generation of Near-Target Adversarial Examples for Alignment Training
- Token Buncher : Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning
- Detecting Data Poisoning in Code Generation LLMs via Black-Box, Vulnerability-Oriented Scanning
Systems & Software Security · Rising
60 papers this week · average of the previous 4 weeks: 12.0
The count rises to 60 against a 12 average: much of this is a large conference's proceedings landing, so read the jump with caution. The content, though, is strong on defeating memory protections: a fuzzer breaks the V8 JavaScript engine sandbox, other work bypasses ARM kernel asynchronous MTE with ordinary corruption bugs, and an LLM agent generates proof-of-vulnerability inputs in minutes. For insurers, defenses presented as robust fall to now-public techniques.
Papers cited:
- TANTRUM: Breaking the Heap Sandbox in JavaScript Engines via Protection Domain Fuzzing
- MTEscape: Bypassing Asynchronous Kernel MTE via Conventional Memory Corruption Vulnerabilities
- PBFuzz: Agentic Directed Fuzzing for PoV Generation
- Insecurity Through Obscurity: Veiled Vulnerabilities in Closed-Source Contracts
Post-Quantum Cryptography · Rising
57 papers this week · average of the previous 4 weeks: 23.8
Post-quantum papers rise to 57 against a 23.8 average, partly from conference proceedings. The most concrete thread is implementations: a non-profiled attack recovers the ML-DSA key with few traces, another shows that Falcon's square-root computations create vulnerabilities. Threshold signatures and post-quantum key recovery also appear. For insurers, the security of standard schemes depends on how they are implemented on devices.
Papers cited:
- When Module Lattice Leaks: Horizontal Fusion Attacks on ML-DSA Implementation
- Square Root of All Evil: The Dangers of Falcon's Superfluous Square Roots
- BeaverDam: A Post-Quantum Threshold Key Recovery Protocol from a One-Time Threshold Oblivious Random Function
- Practical Post-Quantum Quorum Privacy via Structured Threshold Ring Signatures
Hardware & Side Channels · Rising
44 papers this week · average of the previous 4 weeks: 8.2
The count rises to 44 against an 8.2 average, lifted by conference proceedings. The work looks at side channels on implementations: a non-profiled analysis recovers the ML-DSA key with few traces, another refines compression side-channel techniques, a third fingerprints Wasm applications inside a TEE from file-system interactions. For insurers, security depends on how standard schemes are implemented, not only on their choice.
Papers cited:
- When Module Lattice Leaks: Horizontal Fusion Attacks on ML-DSA Implementation
- Criminology: Refined Techniques for Compression Side-Channel Attacks
- POSTER: Fingerprinting WebAssembly Applications in Trusted Execution Environments via File-System Interactions
Identity & Authentication · Rising
52 papers this week · average of the previous 4 weeks: 10.2
The count rises to 52 against a 10.2 average, largely from conference proceedings. The work looks at real-world authentication: new OAuth login abuses in super-app mini-programs, detection of firmware authentication bypass from initial credentials, and password inference from smart-glasses video. For insurers, access controls taken as given can be bypassed in ways standard questionnaires miss.
Papers cited:
- Mini-Programs, Mega-Problems: Unveiling OAuth-based Authentication Misuses in Mini-Programs via Dynamic Analysis
- FirmCred: Detecting Authentication Bypass Vulnerabilities in Firmware from Credential Initialization Perspective
- Watch and Crack: Password Inference from Smart-Glasses Video
How this digest is made
New papers are collected every day and selected every week from arXiv, the IACR ePrint archive and the main security conferences and journals (IEEE S&P, USENIX Security, ACM CCS, NDSS, CRYPTO, PoPETs, IEEE TIFS and others). A score based on source, topic and early signals picks the candidates; an AI model assesses them and writes the summaries. Summaries are based on the abstracts: the original paper is the reference for details.
New papers assessed for this edition: 920.