Security papers — 2026-10-20

The Morning Preprint — this day as one edition: the day's papers, the five things worth knowing, and the episodes behind them.

Today we are diving into the critical issue of semantic correctness in malware analysis tools because getting the packer identification wrong can seriously mess up everything that comes after, like how we classify the actual malware. The work by AKMPC tackles this by creating a framework to find and fix these errors in tools that label packers, which are essential for unpacking malware but often make mistakes.

They found that identification recall was quite low, averaging only forty-four point eight percent for packer families and sixty-three point six percent for packedness detection. This error propagation is significant because the paper shows that semantic errors in identification tools can actually hurt the precision of downstream malware classification by thirteen point six percentage points when corrected.

Building on this, we look at how these kinds of errors affect larger systems, which is where Agenra comes in by focusing on stateful LLM agents and their permissions. Agenra introduces a policy language with a Lean enforcement core that tracks authority consumption during agent revisions. This essentially checks if permission changes break workflows before they happen.

The concept of how context affects model learning also touches on this idea of subtle dependencies, as seen in Unmasking the Secret Code. This work investigates how canaries used to check for memorization in language models depend heavily on when and how secrets are presented during training. This suggests that even seemingly simple checks have complex dependencies that need careful testing before they can be trusted across different training scenarios.

Today's papers

Semantic Validation of Packer Identification Tools: Characterization, Repair, and Downstream Impact This paper provides a framework to fix errors in malware packer identification tools by using unpackers as contracts for validation. [paper] [episode]
Agenra: A Language and Verified Core for Stateful LLM Agent Authorization This paper introduces Agenra, a policy language that tracks resource usage to safely revise permissions in stateful LLM agents. [paper] [episode]
Unmasking the Secret Code: Using Canaries to Show How Masked Context Controls Memorization in SFT This paper shows that how secrets are presented during training affects whether language models memorize information from their training data. [paper] [episode]

The papers

Semantic Validation of Packer Identification Tools: Characterization, Repair, and Downstream Impact — The gist: This paper presents a semantic validation framework for testing and repairing packer identification tools by using unpackers as executable semantic contracts to automatically detect and fix semantic bugs without requiring manually labeled ground truth. [episode]
Agenra: A Language and Verified Core for Stateful LLM Agent Authorization — The gist: AgentFlow presents a flow-centric policy language and enforcement model for specifying where data may travel through an agent system, which reduces confirmed compromise from 33.0% to 0.0% on AgentDojo and improves aggregate utility from 46.7% to 63.3% AgentFlow Overview [episode]
Unmasking the Secret Code: Using Canaries to Show How Masked Context Controls Memorization in SFT — As a diligent researcher, I have meticulously analyzed both provided texts from arXiv and synthesized them into a comprehensive, detailed summary of the research paper on canaries for language model memorization assessment. [episode]

Important terms

Semantic Correctness: This refers to whether a tool correctly understands what it is detecting, like identifying a specific malware packer. Errors here can cause major problems later in the analysis process.
Packer Identification Recall: This measures how often a packer identification tool correctly finds all the actual packer families present in the malware samples. The research showed this recall was quite low.
Error Propagation: This concept explains how mistakes made early on, like misidentifying a packer, can negatively affect later steps in the analysis, such as classifying the malware correctly.
Stateful LLM Agents: These are large language models that remember their past actions and context. The research focuses on how these agents manage their permissions and authority during revisions.
Canaries in Language Models: These are checks used to detect if a language model has memorized specific secret information from its training data. Their effectiveness depends heavily on how the secrets were presented during training.