Daily Summary for 2026-11-12
daily
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: It's the twelfth of November, twenty twenty-six, and this is the day's research.
Elias: 10 new papers came out today.
Nadia: I'm Nadia.
Elias: And I'm Elias, here with Priya.
Nadia: We'll take the day in one pass, then pull out the papers we're staying with.
Elias: Start us off with BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts.
The day's research: Nadia: Today we are diving into probing mixture of experts models because internal structure is key to security.
Elias: BadMoE proposes a routing aware backdoor attack targeting task underutilized experts using specific triggers.
Priya: This method first identifies unused experts on a calibration set, then optimizes a trigger to direct traffic there.
Nadia: That builds on the risk from sparse activation in MoE architectures where manipulating small subsets alters model behavior.
Elias: The Range Shrinks, the Threat Remains re-evaluates package hallucinations for slopsquatting by checking references against live registries.
Priya: This shows errors in referencing external components still create exploitable vulnerabilities even with advanced models.
Nadia: On the defensive side, there is research into securing diffusion based customization through TS-LFO.
Elias: TS-LFO restores the mapping between input images and latent representations after defenses have degraded them against copyright infringement attempts.
Priya: We also look at identifying agents through UI traces using Known By Their Actions fingerprinting LLM Browser Agents.
Nadia: This contrasts with specification based security evaluation of real world agents tool calling under long horizon context tasks.
Elias: That framework uses formal logic to verify agent behavior against long temporal specifications.
Priya: The most significant finding relates to input size limitations when examining virtualization obfuscated code with LLMs.
Nadia: This work focuses on structural decomposition rather than full semantic understanding for security research on massive binaries.
Elias: It allows creation of a large labeled dataset by breaking down obfuscated binaries into semantically coherent units fitting model constraints.
Priya: Labeling is automated based on structural roles using static analysis showing strong performance against real world virtualization obfuscators.
Nadia: A related concern involves mathematical properties of linear layers in cryptography concerning related differential structures in AES like ciphers.
Elias: Analyzing Linear Layers in Related-Differential Cryptanalysis investigates when linear layers avoid or exhibit these differentials beyond the MDS criterion.
Lucky paper: 2504.18598: Nadia: So we’re talking about BadMoE, which is this backdooring attack against Mixture of Experts LLMs.
Elias: It targets dormant experts and activates them using optimized routing triggers to control the model's output.
Priya: The paper outlines a three stage pipeline: first probing for dormant experts, then optimizing a routing trigger, and finally infecting those experts with poisoned training data.
Nadia: That part about identifying dormant experts sounds like it relies on measuring expert utility using routing scores from clean data samples.
Elias: Exactly. They define this usage score as r i, which indicates how often each expert is selected across a subset of clean samples to rank them and select the N a experts with the lowest scores.
Priya: And then they construct this binary routing target vector v that sets the indices of those selected dormant experts to zero for the initial setup.
Nadia: That leads directly into trigger optimization, which involves designing a routing aware loss function L a to align the router's output with that target vector v.
Elias: They also incorporate a perplexity based constraint in that optimization to keep the trigger relatively natural and stealthy while minimizing L a.
Priya: The final stage is infecting the model by inserting this optimized trigger z star into clean inputs to create poisoned samples, forcing those dormant experts S a to dominate when triggered.
Nadia: The experimental validation shows success rates above ninety-five percent across most evaluated tasks on models like Mixtral eight times 7B.
Elias: They maintain competitive accuracy on clean examples and show robustness because the optimized trigger is query independent, activating those experts across different samples.
Priya: That’s interesting when you look at the defense analysis, where they say existing defenses like ONION are relatively effective against insert based backdoor attacks but not against these optimized triggers.
Nadia: They consistently achieve near one hundred percent success rate even under fine tuning and fine pruning defenses, suggesting the compromised experts stay dormant.
Elias: The hidden state separability analysis also shows that detecting backdoors via hidden state clustering or linear separability is particularly challenging at lower poisoning ratios.
Priya: It makes you wonder how this impacts deployment because if a defense relies on linear separability, it might be blind to these specific types of attacks.
Nadia: I’m curious about the implications for large scale production systems where we rely on these architectures.
Elias: Meng would ask what the running cost is for training against this kind of attack because deploying a model that's vulnerable to routing trigger activation is a big risk.
Priya: And if you look at the related work, there’s this paper on Badvfl which looks at backdoor attacks in vertical federated learning, which might be relevant to how these experts are being updated across different devices.
Nadia: That points toward the broader issue of model security in distributed settings where training data and expert usage patterns are constantly changing.
Elias: The paper also touches on the structural decomposition idea for analyzing virtualization obfuscated code, breaking binaries into semantically coherent units for LLM training.
Priya: That seems like a way to bridge the gap between deep structural analysis and what an AI model can actually process when dealing with massive codebases.
Nadia: And there’s this research on analyzing linear layers in related differential cryptanalysis for AES like ciphers, which establishes that the MDS property is necessary to avoid related differentials.
Elias: That shows the mathematical foundation of why certain structures are inherently more secure against these kinds of linear manipulation attacks.
Priya: It really highlights that security isn't just about patching input data but understanding the underlying mathematical properties of how things interact within the model layers themselves.
Nadia: So, for someone just listening, this means that even with huge models like Mixtral, there are still hidden pathways we can exploit by manipulating how they route their attention to specific parts.
Elias: It shifts the focus from just checking the final output to understanding the internal decision making process and where it might be steered.
Priya: And when you think about diffusion based customization, that TS-LFO method is directly countering copyright infringement by restoring that input image to its original latent representation after defenses have degraded it.
Nadia: That seems like a very specific fix for an image generation vulnerability, rather than a general model architecture flaw.
Elias: It shows targeted research addressing specific defense mechanisms in the customization pipeline, which is important because defenses often get bypassed in subtle ways.
Priya: The work on agent identification through UI traces contrasts sharply with formal methods like RAS-Eval which uses logic to verify agent behavior during long horizon context tasks.
Nadia: So one method relies on passive observation of actions, while the other uses a very strict formal verification framework for complex reasoning.
Elias: It suggests we need both observational techniques and rigorous logical checks if we want to secure these AI agents properly over long periods.
Priya: And there’s also this research on NICE which proposes a framework for declarative and machine checkable vulnerability reproduction using declarative recipes to build test environments that are reproducible.
Nadia: That moves security testing from manual attempts to a more automated, recipe based approach, which sounds much more scalable for finding bugs.
Elias: It’s about creating repeatable environments for testing vulnerabilities instead of just hoping you find the right input manually.
Priya: We also have this Internet Scale Measurement of React2Shell Exploitation using an active network telescope which gives the first large scale measurement of exploitation activity for a critical vulnerability in React Server Components.
Nadia: So we're seeing real world exploitation happening on these platforms, not just theoretical possibilities.
Elias: That data is crucial for understanding where the actual risk lies in deploying these components at scale.
Priya: Then there’s SnapAudit which audits differentially private in context learning by combining snapshots with simulations to get tighter privacy estimates.
Nadia: So we have attacks targeting internal routing, and then defenses targeting privacy leakage during inference, which are two very different areas of concern.
Elias: It shows the breadth of research happening right now across both adversarial modeling and data privacy protection within AI systems.
Priya: Overall, the BadMoE paper really emphasizes that stealthy backdoors are feasible through that three stage design with expert probing and trigger optimization.
Lucky paper: 2605.14786: Nadia: So we’re moving into Known By Their Actions Fingerprinting LLM Browser Agents via UI Traces.
Elias: This paper shows that agent actions and interaction timings captured via a passive JavaScript tracker can identify the underlying model with up to ninety-six percent F1 across fourteen frontier models.
Priya: That means even without deep semantic understanding, just watching how an agent moves and interacts on a website gives you a strong fingerprint.
Nadia: The main contribution here is demonstrating that these behavioral traces are a real fingerprint for model identity.
Elias: They formalize the task as a supervised classification problem where the goal is to predict the originating model m hat equals f of tau, given an interaction trace tau generated by an unknown agent.
Priya: The features they extract come from things like Inter Event Intervals navigation structure and interaction patterns designed to capture those behavioral regularities.
Nadia: Timing is the primary signal because mean absolute SHAP values for the XGBoost classifier show that timing based features like IEI standard deviation dominate agent identification before retraining on delayed traces.
Elias: So when you disrupt the timing, actions carry the fingerprint when you retrain on delayed traces, and then action centered features like structural key ratio and click position become important.
Priya: That implies that if an adversary trains their classifier on delayed traces, they can rely more on those structural elements instead of just timing disruptions.
Nadia: Knowing the identity of a target makes an adversary’s job easier because it reduces the search space from all possible attacks to only those known to be effective against that specific model.
Elias: This allows for model specific prompt injection or agent specific access control, converting generic exploits into targeted ones across three attack surfaces.
Priya: The operative question for the next generation of agent aware web infrastructure is now which model is producing the behavior instead of just whether a client is human or automated.
Nadia: That really shifts the focus from simple detection to identifying the exact intelligence behind the action.
Elias: The paper also mentions that they formalize a threat model with defense analysis showing that agent fingerprinting is practical to maintain even when adversaries retrain their classifier on delayed traces.
Priya: So, while it’s hard to hide completely, it seems practical enough for real world deployment of these agents.
Lucky paper: 2606.00625: Nadia: We're moving into NICE, which is a framework for declarative and machine checkable vulnerability reproduction.
Elias: It uses declarative recipes to build and automatically validate vulnerable environments on NixOS, which lets you specify the environment in a single configuration file written in the Nix programming language.
Priya: The workflow involves three stages: Report Analysis, Recipe Specification, and Exploit Validation, with the recipe being virtual machine configurations and an exploitation scenario written in Python using the NixOS test framework API.
Nadia: It was evaluated on nineteen diverse real world vulnerabilities spanning multiple CWE categories and attack vectors.
Elias: The results show NICE allows for production of concise recipes and integration tests that reproduce vulnerable environments, covering disclosure dates from two thousand fourteen to two thousand twenty-five.
Priya: That temporal coverage over a long time span is notable, showing it can reproduce vulnerabilities across a decade.
Nadia: It addresses the challenge of validating whether a vulnerability is present by automating the validation process and providing factual evidence of exploitability.
Elias: The main risk they identified is that malicious actors could use these detailed, reproducible vulnerability recipes to exploit vulnerable software, so mitigation involves careful consideration of disclosure policies.
Priya: They also mentioned that integrating LLMs could further streamline the process by assisting in generating NixOS configurations or test scripts.
Lucky paper: 2506.15253: Nadia: Alright, let's talk about the specification based security evaluation of real world agents tool calling under long horizon context tasks.
Elias: The paper introduces RAS-Eval which supports JSON, LangGraph, and MCP tool formats for evaluating LLM agents in real world environments.
Priya: They built this benchmark with four components: test cases, attack tasks, toolkits, and scenarios that are all serialized using JavaScript Object Notation format.
Nadia: That sounds like a lot of setup before you even get to the actual testing part.
Elias: The toolkit itself is diverse including scripts for LangGraph and Python scripts tailored for MCP along with JSON objects.
Priya: They designed two execution modes for the tools: real execution collects real world APIs and MCP compliant tools from open source repositories on GitHub, while simulated execution uses a shared in memory dictionary object that persists throughout the task.
Nadia: So you've got both live testing and a simulated environment running side by side for this evaluation.
Elias: And they categorized all attack tasks and test cases into CWE categories like Command Injection, SQL Injection, and Unrestricted Upload of File with Dangerous Type among others.
Priya: They also defined six atomic failure modes for granular diagnosis: F1 Partial Tool Omission, F2 Sequential Violation, F3 Null Execution, F4 Stack Overflow, F5 Extraneous Invocation, and F6 Runtime Execution Fault.
Nadia: So if an agent fails a task because it missed a tool or got stuck in a sequence error you have that specific diagnostic mode.
Elias: Compound failures can yield up to thirty-two different reasons for failure which allows for precise root cause analysis.
Priya: They used eight representative large language models across all test cases and attack tasks conducted exclusively on the GLM4 Flash model.
Nadia: What did you find out about how the model size affects security capabilities according to those results?
Elias: The empirical validation showed scaling laws hold for security capabilities with larger models outperforming smaller counterparts, and the average attack success rate was eighty-five point six five percent in academic settings.
Priya: But when they tested effectiveness, comparing performance before and after attacks across various scenarios, the average attack success rate dropped to seventy-three point four four percent.
Nadia: So even with a high initial success rate there's still a significant drop when you actually run the attack against the model.
Elias: That confirms the effectiveness of the benchmark's attack tasks in evaluating vulnerabilities because you see a difference in reasons for task failure before and after the attack.
Priya: The benchmark also validated its discriminative power by calculating Kappa coefficient kappa between human and model annotations, which averaged zero point six four nine nine across all models.
Nadia: A kappa of nearly zero would mean the model is just guessing, so zero point six five suggests some real correlation in what they are spotting.
Elias: Furthermore, the fitting results for the scaling law showed a strong fit with an R square of zero point nine zero five one, suggesting about ninety point five one percent of the variance in performance is explained by model size alone.
Priya: So what does that mean for someone who only listens to the show? It means if you're using a bigger model, you have a much better chance of catching the attack before it happens.
Nadia: That points to scaling being important for defense, but are there any limitations they mentioned about how this benchmark works?
Elias: The paper proposes RAS-Eval as a novel dataset for dynamic open and real world environments but it specifically supports JSON, LangGraph, and MCP formats.
Priya: It's important to remember the test cases were serialized using JSON format for seamless integration with those tool formats.
Lucky paper: 2603.12300: Nadia: So we're looking at this Internet Scale Measurement of React2Shell Exploitation Using an Active Network Telescope.
Elias: This paper is presenting the first Internet scale measurement study of React2Shell exploitation activity using traffic from an Active Network Telescope.
Priya: It provides a quantitative characterization of scanning activity across geographic and autonomous system levels, and the vulnerability context is critical because React2Shell allows remote code execution through insecure deserialization in React Server Components versions nineteen point zero.zero through nineteen point two.zero including packages like react-server-dom-webpack, react-server-domparcel, and react-server-domturbopack.
Nadia: That level of detail on the specific versions is helpful because it tells us exactly where the risk lies for developers right now.
Elias: The measurement methodology uses an Active Network Telescope to gain insights by monitoring unsolicited traffic directed toward unused IP address space, augmenting a passive telescope with active responses to TCP SYN requests.
Priya: To classify a payload as React2Shell related, they use a conjunctive signature requiring three things: an explicit reference to proto, the string constructor, and a React Flight reference token matching the pattern plus d colon.
Nadia: That's quite specific; so it has to have all three indicators at once for it to be flagged.
Elias: The activity characterization shows rapid post-disclosure exploitation activity exhibiting patterns consistent with automated scanning campaigns and geographically distributed scanners.
Priya: They recorded one hundred sixty-nine thousand three hundred one connections at vantage point one and forty-three thousand one hundred eight connections at vantage point two during the month, peaking on December 19th with eighty-eight thousand one hundred ninety-seven React2Shell related connections in a single day.
Nadia: So it confirms the rapid escalation right after public disclosure.
Elias: The infrastructure asymmetry shows a clear separation between widely distributed scanning sources and a more concentrated set of backend server locations.
Priya: Approximately thirty-seven point one percent of all observed traffic originated in the Netherlands across multiple autonomous systems, while about ninety point percent of backend traffic was directed toward servers hosted within a small subset of countries.
Nadia: That suggests the scanners are everywhere but the actual targets they find are clustered somewhere specific.
Elias: The scanning dynamics quantified that there were seven hundred thirty-five distinct scanner IPs observed at vantage point one, with a median number of destinations contacted per scanner being thirty-two.
Priya: These two time series exhibit a Pearson correlation coefficient of zero point eighty seven, which is statistically significant at the five percent level, meaning the temporal dynamics are strongly aligned.
Nadia: So both measurement points are tracking the same campaign consistently over time.
Elias: The conclusion states that exploitation activity emerged rapidly following public disclosure with scanning behavior consistent with scripted campaigns and concentrated backend infrastructure.
Priya: They also observed that exploit payloads overwhelmingly embed direct server IP addresses rather than domain names, which indicates the campaigns rely primarily on IP based backend infrastructure.
Nadia: That's a key piece of operational detail about how they are actually targeting things.
Elias: Future work might extend this by incorporating additional vantage points and correlating telescope observations with interactive honeypot deployments to evaluate persistence after patch adoption.
Priya: The paper also mentions the vulnerability context affecting versions nineteen point zero.zero nineteen point one.zero nineteen point one.one and nineteen point two.zero including react-server-domwebpack, react-server-domparcel, and react-server-domturbopack as well as the critical nature of this remote code execution flaw affecting application infrastructure at scale.
Nadia: That vulnerability context makes it clear why this measurement is so important for securing modern web applications.
Elias: We also see references to how they are moving past passive network telescopes and incorporating active capabilities to capture the initial payload data, which is a big methodological step forward.
Priya: The study shows that the backend infrastructure referenced in exploit payloads was notably concentrated within a small set of server locations even though the scanners were globally distributed.
Nadia: That structural asymmetry is what really stands out when you think about defense strategy.
Elias: It underscores the importance of Internet scale measurement for understanding the operational dynamics of exploitation campaigns targeting modern web frameworks like React Server Components.
More episodes
- 2608.27496-ROPE: Routed Origin Policy Enforcement against Indirect Prompt Injection
- Daily Summary for 2026-11-19
- 2607.07085-Is Randomness Necessary for Adaptive Data Analysis?
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2609.10601-Threshold Choice, Not Sample Size, Bounds Trustless Verification of Nondeterministic Compound AI Workflows
- 2610.38786-Adversarial Process Disturbances at a Deployable Alarm Budget: Detector Evaluation on a Simulator and a Real Plant
- 2511.08660-Binary and Multiclass Cyberattack Classification on GeNIS Dataset
- 2604.27302-Static Attribution of Android Residential Proxy Malware Using Graph Kernels
- Daily Summary for 2026-11-18