RESEARCH
January 2026: signed drivers as weapons, safety neurons, password managers
RansomwareSupply chainOT and industrial security
January 2026 is NDSS month. Malware brings its own signed, vulnerable drivers to switch off defences, and an architecture like ACE separates agent planning from untrusted apps. A handful of neurons hold up LLM alignment, desktop password managers give way to simulated keystrokes and email aliases confuse identities. IoT firmware, PLCs and PyPI packages round out the month.
Archive: reconstruction written in November 2026
1. Unveiling BYOVD Threats: Malware's Use and Abuse of Kernel Drivers
Network and Distributed System Security Symposium · Andrea Monzani, Antonio Parata, Andrea Oliveri, Simone Aonzo et al. · Malware & Threat Intelligence
In short: A sandbox that follows drivers deep into the kernel exposes BYOVD attacks and seven unknown vulnerable drivers.
In BYOVD attacks, malware loads a legitimate, signed but vulnerable Windows driver to enter the kernel and switch off defences. Public sandboxes see only the load. The authors build a taxonomy of stages and a virtualisation-based sandbox that follows every step of the driver: across 8,779 samples and 773 drivers they flag 48 suspicious cases and seven previously unknown vulnerable drivers.
Why it matters. Ransomware and state groups use the technique to disable EDR: controlling which drivers can load cuts a vector behind severe claims.
2. Actively Understanding the Dynamics and Risks of the Threat Intelligence Ecosystem
Network and Distributed System Security Symposium · Tillson Galloway, Omar Alrawi, Allen Chang, Athanasios Avgetidis et al. · Malware & Threat Intelligence
In short: By tracking watermarked samples through the threat-intelligence chain, the authors find delays, bottlenecks and exploitable gaps.
The threat-intelligence ecosystem moves billions, yet nobody knew how it really works. The authors submit samples with watermarked network indicators and follow their spread among vendors. When information flows, threats are almost always disrupted, but vendors that share selectively or hours to days late slow everyone, and easy-to-fingerprint sandboxes are already being evaded.
Why it matters. How fast an indicator reaches your tools depends on the vendor chain: worth weighing when choosing providers.
3. DySec: A Machine Learning-Based Dynamic Analysis for Detecting Malicious Packages in PyPI Ecosystem
IEEE Transactions on Information Forensics and Security · Sk. Tanzir Mehedi, Chadni Islam, Gowri Sankar Ramachandran, Raja Jurdak · Software Supply Chain
In short: DySec watches a PyPI package with eBPF as it installs and spots malicious ones with 96% accuracy.
Malicious Python packages exploit trust in open-source repositories through look-alike names and payloads generated on the fly. DySec monitors installation with eBPF probes and measures 36 behaviours, from system calls to network traffic. On 14,271 packages it reaches 96% accuracy and found six malicious packages that PyPI had classed as benign.
Why it matters. A malicious package installed by one developer can reach many products: install-time behaviour analysis closes a gap.
4. ACE: A Security Architecture for LLM-Integrated App Systems
Network and Distributed System Security Symposium · Enlai Li, Tushin Mallick, Evan Rose, William Robertson et al. · AI & LLM Security
In short: ACE plans an agent's work from trusted information only and isolates third-party apps during execution.
In systems where an LLM invokes third-party apps, a malicious app can tamper with the plan, stall execution or steal data. The authors show new attacks, including against IsolateGPT, and propose ACE: first an abstract plan built only from trusted data and checked against user constraints, then execution with barriers between apps. It withstands indirect-injection benchmark attacks.
Why it matters. Applying classic systems-security principles to agents is the path to making them controllable, and so assessable.
5. NeuroStrike: Neuron-Level Attacks on Aligned LLMs
Network and Distributed System Security Symposium · Lichao Wu, Sasha Behrouzi, Mohamadreza Rostami, Maximilian Thang et al. · AI & LLM Security
In short: By switching off under 0.6% of neurons in some layers, NeuroStrike bypasses the alignment of over 20 open LLMs.
LLM safety alignment relies on a few neurons specialised in spotting harmful requests. With model access, NeuroStrike finds and disables them: removing under 0.6% of neurons yields a 76.9% average success rate across more than 20 open models. Without access, it trains prompt generators on similar models and reaches 63.7% against five closed models.
Why it matters. Safeguards in open-weight models come off with little effort: users must add their own downstream controls.
6. Vault Raider: Stealthy UI-based Attacks Against Password Managers in Desktop Environments
Network and Distributed System Security Symposium · Andrea Infantino, Mir Masood Ali, Kostas Solomos, Jason Polakis · Identity & Authentication
In short: On Windows and Linux, a malicious app can steal passwords and codes from password managers via invisible simulated keystrokes.
In the browser, password-manager autofill resists phishing because it knows the domain. On the desktop, with native apps, things change. The authors analyse popular managers such as 1Password and LastPass on macOS, Windows and Linux: macOS offers good protection, while Windows lacks checks and another app can steal credentials, one-time codes and vault secret keys.
Why it matters. The password manager concentrates every credential: a weakness there magnifies the impact of an infection.
7. One Email, Many Faces: A Deep Dive into Identity Confusion in Email Aliases
Network and Distributed System Security Symposium · M. Wu, Geng Hong, Jiatao Chen, Baojun Liu et al. · Identity & Authentication
In short: Email aliases confuse providers and platforms: 139 npm accounts from one address, and users more prone to phishing.
Providers treat aliases such as [email protected] as entries to the same mailbox, while many platforms see them as separate identities. Across 28 providers and 18 platforms, only Gmail fully documents its rules and 11 providers support undocumented aliases. An attacker created 139 npm accounts from one email, and 31.65% of users who know aliases mistake phishing emails for legitimate ones.
Why it matters. Email is the recovery key for almost every account: ambiguous alias rules open the way to fraud and identity theft.
8. FirmAgent: Leveraging Fuzzing to Assist LLM Agents with IoT Firmware Vulnerability Discovery
Network and Distributed System Security Symposium · Jiangan Ji, Chao Zhang, Shuitao Gan, Lin Jian et al. · CPS & IoT Security
In short: FirmAgent combines fuzzing and LLM agents and finds 140 new vulnerabilities in 14 real IoT firmware images.
Static analysis of IoT firmware, LLM-based or not, yields many false positives, while fuzzing misses many bugs. FirmAgent uses fuzzing to find where input enters and rebuild risky paths, then one LLM agent analyses those paths and another prepares test cases. On 14 real firmware images it finds 182 vulnerabilities with 91% precision, 140 of them new and 17 with CVEs.
Why it matters. IoT devices on corporate networks carry many hard-to-see flaws: better tools speed up discovery, for defenders and attackers alike.
9. Discovering Blind-Trust Vulnerabilities in PLC Binaries via State Machine Recovery
Network and Distributed System Security Symposium · Fangzhou Dong, Arvind S. Raj, Efrén López-Morales, Siyu Liu et al. · CPS & IoT Security
In short: Ta'veren rebuilds a PLC's state machine from its binary and finds missing safety checks on inputs.
PLCs control plants with physical effects, and a safety flaw can have severe consequences. Many vulnerabilities stem from missing or incomplete checks on sensor inputs, which the PLC trusts blindly. Ta'veren finds these flaws directly from binaries, with no source code or emulation, by recovering the controller's state machine, and uncovers critical safety violations in real PLCs.
Why it matters. In industry a logic defect can cause physical damage and downtime: analysis that needs no vendor code helps plant operators.
Trends of the month
AI & LLM Security
525 papers this month on this topic
With 525 of 2,920 papers, AI security remains the top topic. NeuroStrike shows alignment rests on very few neurons, easy to switch off in open models, while ACE applies classic systems-security principles to agents, separating planning from untrusted data. Protection is moving from the model to the architecture around it.
Papers cited:
- ACE: A Security Architecture for LLM-Integrated App Systems
- NeuroStrike: Neuron-Level Attacks on Aligned LLMs
Malware & Threat Intelligence
79 papers this month on this topic
Three papers look at where malware hides and how quickly it is stopped. Signed vulnerable drivers remain an effective way to switch off EDR, the threat-intelligence chain works but has bottlenecks of hours or days, and watching install-time behaviour catches malicious PyPI packages that static checks miss.
Papers cited:
- Unveiling BYOVD Threats: Malware's Use and Abuse of Kernel Drivers
- Actively Understanding the Dynamics and Risks of the Threat Intelligence Ecosystem
- DySec: A Machine Learning-Based Dynamic Analysis for Detecting Malicious Packages in PyPI Ecosystem
Identity & Authentication
143 papers this month on this topic
Digital identity shows unexpected weak spots. Password managers, safe in the browser, can hand credentials and codes to another app on the desktop, and unclear email-alias rules let attackers multiply accounts and make phishing more convincing precisely for experienced users.
Papers cited:
- Vault Raider: Stealthy UI-based Attacks Against Password Managers in Desktop Environments
- One Email, Many Faces: A Deep Dive into Identity Confusion in Email Aliases
How this digest is made
This issue belongs to the digest archive, reconstructed after the fact. The month's papers were collected from arXiv, the IACR ePrint archive and the main security conferences and journals; the selection also weighs the citations received since publication. An AI model assessed the candidates and wrote the summaries, which are based on the abstracts: the original paper is the reference for details.
Papers of the month assessed: 2920.