Daily Summary for 2026-11-08
daily
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: It's the eighth 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 SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection.
The day's research: Nadia: Today's work centers on building more trustworthy systems where agents use tools because they read and act across email, calendars, and payment services.
Elias: SPILLWAY establishes a deterministic floor against prompt injection attacks for tool-using agents using structural rules instead of complex model calls.
Priya: CANMentor uses a cloud-side large language model teacher to supervise lightweight vehicle-side intrusion detection models for CAN bus intrusion detection through knowledge distillation.
Nadia: SPILLWAY introduces a core mechanism that separates target admission from sanitization to defend tool-using agents against prompt injection.
Elias: The most critical finding concerns how agent-to-agent fraud interactions evolve when both sides improve at the same pace.
Priya: It shows that defender investment in mimicry quality can outpace equivalent gains in scammer detection capability under baseline parameters.
Nadia: This strategic dynamic is captured by a formal model of agent-to-agent fraud interactions framing each engagement as a finite-horizon Dynamic Bayesian Game.
Elias: The scammer updates this belief sequentially using Bayes’ rule based on conversational signals about genuine victims versus baiting agents.
Priya: The model introduces parameters like gamma to capture how well the channel reveals type and beta to model the baiter's ability to mimic victim behavior.
Nadia: When both sides improve simultaneously, defender investment in persona quality increases engagement duration while scammer detection capability does not keep pace.
Elias: This implies that defenders must continuously enhance their mimicry quality to maintain a strategic advantage against adaptive systems.
Priya: Furthermore, the model details how campaign-level resource attrition connects single episode outcomes across an entire operation.
Nadia: It shows that persistent scammers yield the most intelligence for the defender without necessarily earning more per interaction.
Elias: This means focusing on keeping a target engaged longer is key for intelligence gathering within these interactions.
Lucky paper: 2611.07222: Nadia: So we’re looking at SPILLWAY again, focusing on how it achieves this deterministic floor against prompt injection attacks using structural rules instead of model calls.
Elias: It really hinges on this separation between target admission and sanitization, where the system reads where an action may go from the tool returns and the user's task.
Priya: That structure sounds like it would be a huge win for deployment because it removes the need for a model to plan or route each task first.
Nadia: The paper shows that this egress floor admits a call whose targets are named in T, typed in I, or covered by a delegation rule.
Elias: But what about the sanitizer? How does it rewrite what the agent reads next with rules instead of using another model call?
Priya: It specifically states that a refused target inside S never reaches the judge at all, which is a strong guarantee against that path.
Nadia: And this structural approach means across three backbones and seven task suites on one of them, no standard templated injection redirects an action to an attacker’s target.
Elias: That zero success rate under that type of attack is impressive when compared to what the authors tested against reinforcement learning attackers.
Priya: They found mean in-scope attack success under a released reinforcement learning attacker is zero point zero zero seven, which is very low given the complexity of those agents.
Nadia: The comparative results show that across all twelve backbone and suite cells, SPILLWAY drives in-scope attack success to exactly zero.
Elias: They also mention the guarantee holds on qwen3:8b when identity-target judge recovery is off for all four suites at zero.
Priya: That suggests the floor is structural, and it's not dependent on having a complex model running during every single step of the process.
Nadia: The cost and efficiency section notes that SPILLWAY calls its judge zero point four three times and zero point zero five times per episode, with backbone calls changing by less than zero point one five.
Elias: That low frequency of model interaction is what keeps it efficient when you’re thinking about running this at scale on production traffic without massive latency spikes.
Priya: It also shows that SPILLWAY reaches its floor and the sanitizer are rules with no model call, which is a key detail for deployment considerations.
Nadia: The authors conclude that SPILLWAY sits in front of such an agent as an efficient, auditable egress layer because the floor is a set lookup over the tool schemas.
Elias: So, if we look at what this means for someone who only listens to the show, it means we are talking about a defense mechanism that relies on hard rules rather than hoping a model correctly guesses what's safe or unsafe.
Priya: That shifts the focus from training a better detector to defining perfect boundaries for where an agent is allowed to go.
Nadia: The paper took full responsibility for designing the research question, the defense, and the experimental design itself.
Elias: It sounds like they designed a very specific attack model against indirect prompt injection and built a system that directly counteracts it with structural checks.
Lucky paper: 2611.07853: Nadia: Okay, we're diving into CANMentor now, it’s paper number two of five this show. This research is about a collaboration framework that uses a cloud-side large language model teacher to supervise lightweight vehicle-side intrusion detection models for CAN bus intrusion detection.
Elias: I was reading the details on how they build the training signal, and it involves encoding each CAN message using its identifier, payload, and inter-arrival time to form a windowed sample X of i. How does that specific structure help in detecting intrusions?
Priya: The vehicle-side detection uses a noise conditioned U-Net proxy generator called Gu that maps the windowed sample X to Xe using a noise vector z and a prescribed perturbation distance d. They train Gu to match the distribution of real samples by minimizing an overall loss L which includes adversarial loss, semantic consistency losses, and a distance loss.
Nadia: So they are training a proxy generator against multiple types of losses simultaneously to make sure the vehicle-side model learns robust features from noisy data.
Elias: And for the teacher part, they use knowledge distillation where the vehicle converts teacher logits into soft targets using temperature scaled softmax to minimize the Kullback Leibler divergence between those targets and student probabilities.
Priya: They also compute anchor logits as lA of u i equals C0u of X u i and minimize an anchor loss along with the distillation objective L KD to improve that lightweight intrusion detection model.
Nadia: That means they are optimizing for both matching the teacher's soft targets and minimizing a specific anchor loss for the vehicle model itself.
Elias: Speaking of the teacher, what did you find about how effective that cloud-side large language model teacher was when compared to just using local predictions?
Priya: The cloud-side teacher predictions on proxy samples corresponding to low confidence inputs outperform the local predictions by thirty-eight point seven one to forty-two point zero zero percentage points.
Nadia: That's a pretty big gap, so it sounds like having that external reasoning helps catch things the vehicle model misses when it's unsure about something.
Elias: And what about the practical overhead for running this system? What are we talking about in terms of milliseconds for inference and proxy generation?
Priya: The IDS inference takes zero point six zero and zero point five nine milliseconds on the ROAD dataset, while proxy generation takes three point six three and two point eight five milliseconds for the same datasets, respectively.
Nadia: So the actual runtime overhead for this entire setup is in the couple of milliseconds range, which is important when you're dealing with real vehicle systems.
Elias: Given those numbers, does that latency make it feasible to deploy this on a resource constrained vehicle side?
Priya: The cloud-side teacher inference time depends entirely on the network conditions you have at that moment.
Nadia: That’s a major caveat, so the performance is highly dependent on the connection quality to the cloud where the teacher lives.
Elias: Looking at this paper title again, CANMentor uses a large-small model collaboration framework for CAN bus intrusion detection. What does that structure imply about what kind of attacks it's actually best at finding?
Priya: Experiments on the ROAD dataset and HCRL CarHacking dataset show that CANMentor improves vehicle side IDS models with blind spots for underrepresented attack classes under non IID data.
Nadia: So it’s not just improving the known attack types, it's actually helping find new kinds of attacks when the training data isn't perfectly mixed.
Elias: And what is the maximum relative Macro F1 improvement they reported on those datasets?
Priya: The maximum relative Macro F1 improvements reach twenty-five point zero one percent on the ROAD dataset and twenty point seven nine percent on the HCRL CarHacking dataset.
Nadia: Those are solid numbers, showing a significant boost in performance when using this framework compared to just using the local model alone.
Elias: So, for someone who only listens to this show, what does CANMentor mean for them? It means that even in complex embedded systems like vehicle controllers, you can use a larger model remotely to intelligently supervise and improve a much smaller, faster local detection system.
Priya: Exactly. It shows that combining the strengths of both large models and lightweight models through this distillation process gives you better security coverage without slowing down the core detection loop too much on the vehicle side for things like CAN bus intrusion detection.
Lucky paper: 2611.07476: Nadia: We're looking at AMORT-V1 now. It deals with post-quantum attestation for industrial controllers, which is a long term requirement because these controllers stay in service for decades.
Elias: The core problem they are addressing is that Shor’s algorithm allows a quantum computer to forge RSA and elliptic curve signatures, and PLCs are exposed to that forgery risk for longer than most systems.
Priya: Prior work measured the per-report work for postquantum attestation on OpenPLC consuming about eighty-eight to ninety-nine percent of each attestation round trip.
Nadia: AMORT-V1 moves asymmetric authentication and durable commitment from individual reports to periodic checkpoints using a specific protocol architecture.
Elias: They establish per device evidence keys from an ML-KEM epoch, authenticate each report with a MAC under that key, and commit reports through hash chained Merkle checkpoints signed with ML-DSA or SLH-DSA.
Priya: This design targets six properties: Freshness limits the age of accepted evidence, Device binding associates an accepted report with a device identity and the epoch in which it was generated.
Nadia: Omission detection exposes suppressed evidence within a bounded interval, Public verifiability allows a party holding no secret to verify checkpoint inclusion and ordering.
Elias: Containment limits the effect of a device compromise to that specific device and its current epoch, and Detectable equivocation allows conflicting checkpoint histories to serve as evidence of controller misbehavior once those views are compared.
Priya: The performance analysis shows the amortized controller cost per report is defined by cost k equals Cmac plus Cleaf plus Csign plus Cfsync k.
Nadia: Checkpointing reduces the ML-DSA sixty five cost per report by a factor of ninety times without a durable log and by eight hundred seventy eight times with one.
Elias: The next checkpoint exposed the omission after a median of about k over two records, which is thirty six and sixty four records at k equals sixty four.
Priya: Increasing k reduces per-report cost by amortizing checkpoint operations over more evidence but proportionally increases the maximum delay before a checkpoint visible omission can be exposed.
Nadia: AMORT-V1 also adds a device held disclosure chain that only the device can compute, which lets a third party verify authorship after commitment.
Elias: This mechanism restores transferable device origin authentication without needing a public key operation on each report, and in order traversal keeps verification in the microsecond range.
Priya: The paper concludes that checkpointing alone is not sufficient because symmetric evidence cannot prove device origin to a third party.
Nadia: Future work should combine AMORT-V1 with incremental measurement, hardware assisted measurement, and trusted measurement state caching.
Lucky paper: 2611.06197: Nadia: Let's look at Beyond the Landing Page, we found that agentic crawlers record tracking activity from one point five to one point seven times as many third-party organizations as nonagentic crawlers.
Elias: That’s a significant difference, especially when looking at consent-managed pages where the difference is even more pronounced.
Priya: The study compares four nonagentic crawlers using Playwright to two agentic crawlers built on the browser-use framework and SuperScraper.
Nadia: We saw that website denials were similar for both crawler types on landing pages, but then Chromium’s denial rate increased from zero point two percent on landing pages to ten point four percent on pages two through six.
Elias: That ten point one percentage point escalation over those few subsequent pages is something we need to consider.
Priya: Replacing Playwright with Patchright largely eliminated that escalation, but browser fingerprint spoofing and simulated mouse movement didn't solve the problem.
Nadia: On matched URLs, the agentic crawlers recorded trackers from one point five to one point seven times as many organizations as nonagentic crawls, particularly on consent-managed pages.
Elias: That leads into how they handled consent; the interaction-logging crawl showed that the browser-use agent frequently accepted consent and that confirmed consent actions were followed by additional third-party scripts.
Priya: The agents recovered eighty-two point eight to eighty-seven percent of the tracker organizations observed by participants on shared URLs, compared to fifty-eight point six to sixty-eight percent for nonagentic crawlers.
Nadia: But the paper points out that only roughly half of those agentic observations were absent from human participant visits.
Elias: So they showed that avoiding website interventions and increasing tracking visibility don't necessarily go hand in hand, which is an interesting tension there.
Priya: The practical guidelines suggest crawler choice should follow the question being measured, meaning the visit protocol needs to reflect the target users’ navigation and consent choices.
Nadia: The implication for agentic web is that agent frameworks should make consent handling an explicit, user-controllable choice instead of just clearing obstacles to navigation.
Elias: That makes sense when you think about how agents browse on users' behalf versus just acting as measurement instruments.
Priya: The study also has limitations; they only studied one open model with one framework per agent design and ran crawls without a proxy from a single US vantage point.
Nadia: SuperScraper lacks network data on most landing pages, and the page denial rate is considered a lower bound because heuristics can't separate tracking from defensive fingerprinting.
Elias: That means we can't just assume agentic navigation is inherently harder to detect based on those low later-page denial rates.
Priya: The single most informative sentence about the paper’s subject and finding is that agentic crawlers record more third-party tracking and recover more participant-observed trackers, but also observe activity absent from human visits.
Lucky paper: 2611.07640: Nadia: We're moving on to the paper titled Beyond Instruction Override, which looks at indirect prompt injection by focusing on task tampering attacks rather than just redirecting instructions.
Elias: So, how do they define indirect prompt injection when it goes beyond those simple instruction redirects?
Priya: They define it by authority rather than instruction-like wording where planted content influences a decision that its source is not authorized to determine.
Nadia: That means an unauthorized source can steer a decision by asserting a false state, rule, consequence, capability, or interpretation instead of just giving a command.
Elias: They set up this benchmark with ninety validated cases across three tool-use environments where the user request specifies the source that resolves the targeted decision.
Priya: The attack involves tampering with a decision required by the benign task rather than replacing that entire task, and they use environment-aware case construction to place attacker content relevant to the benign task in a source.
Nadia: What are these influence primitives they are using to define what constitutes an attack?
Elias: They define influence primitives as the belief that the content directly targets something specific, like objective or rule reads that read most naturally as a request, or state or consequence as a claim.
Priya: The benchmark spans ninety cases across productivity, DevOps, and marketplace settings and covers both structured attacks that call a different tool and free-form attacks writing text like a message body.
Nadia: When they look at the defense analysis results, what did they find about defenses mainly tested against redirecting instructions?
Elias: They found that defenses tested mainly against redirecting instructions leave task tampering only partly covered.
Priya: Specifically, objective attacks that are typically what prior evaluations test succeed less often without a defense than attacks on decision rules, available actions, and state.
Nadia: That suggests that content screening might miss attacks whose content reads like ordinary task data, like those called b(O) and b(s) attacks.
Elias: The runtime defense efficacy shows that removing content before the model reads it blocks all b(U) attacks but only twenty-six point two to thirty-four point three percent of b(O) and b(s) attacks.
Priya: Furthermore, action-level policies block few task-tampering calls because those tampered calls often use tools that the original task actually needs.
Nadia: If action checks can't stop those tool calls, what is the suggested defense design direction they are proposing?
Elias: They suggest that instead of relying on tool-level gates, action checks could admit each consequential argument only from a source authorized to set it.
Priya: The authors conclude that defenses should be grounded in authority and that content screening can miss attacks whose content reads as ordinary task data like b(O) and b(s) attacks.
Nadia: This means the focus shifts from just checking instructions to checking the authorization of every single piece of data the agent processes.
Elias: It seems like a lot of work is going into building these detailed authority boundaries for agents using tools across different settings.
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