The day in conversation for 2026-11-02
discussion
Transcript
The discussion: Nadia: The idea of LLMs inferring cognitive biases directly from raw log sequences to map them onto ATT&CK is compelling, but I’m struggling with the operational translation; how do we quantify that inference reliably enough for an adaptive defense system? Elias: From my perspective on the logs side, if we look at the MITRE mapping, it seems like the gap isn't just in mapping events to techniques, but in establishing a reliable baseline for what constitutes a 'cognitive bias' trigger versus standard noise in network traffic. Have you looked into how much variance you need before that inference becomes actionable? Priya: I see the utility here, especially when we consider the Text Detection paper. If we focus on prompt injection defenses, that context-aware classifier handles the malicious instruction itself with some level of adversarial hardening; does that robust training against evasion translate directly to how an LLM interprets a sequence of benign network events as hostile? Nadia: That’s a key question. If the underlying model is trained adversarially, we are protecting the input mechanism, but we still need the log analysis component to distinguish sophisticated lateral movement from routine administrative tasks. What specific metrics are you using to measure that distinction in your text classification model? Elias: We're looking at perplexity scores adjusted for domain specificity; the paper suggests that if you feed it highly specialized IoT telemetry, the model needs a much
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