Veritas: A Semantically Grounded Agentic Framework for Memory Corruption Vulnerability Detection in Binaries

2026-01-20 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

Veritas is a hybrid, semantically grounded agentic framework designed for detecting memory corruption vulnerabilities in stripped binaries. It integrates a static slicer, "defuse", which reconstructs interprocedural value-flow from RetDec-lifted LLVM IR, with a dual-view LLM detector, "discovery". The detector performs step-wise reasoning over statically grounded candidate flows using both RetDec-decompiled C and selectively instantiated LLVM IR. A multi-agent "validator" then confirms or rejects detector hypotheses by grounding them in debugger-visible artifacts and concrete runtime evidence. Evaluated on a curated benchmark, Veritas achieved 90% recall. An exhaustive validation of 623 candidate vulnerabilities produced no false positives. Furthermore, Veritas successfully discovered a previously unknown Apple vulnerability (0day), which was confirmed and assigned a CVE.

Key takeaway

For AI Security Engineers tasked with identifying memory corruption vulnerabilities in stripped binaries, traditional static analysis and ungrounded LLM approaches prove largely ineffective. You should prioritize tools that integrate semantic grounding, like Veritas, which achieved 90% recall and discovered a zero-day. This hybrid approach, combining static flow analysis with runtime validation, significantly reduces false positives and enhances detection reliability, making it crucial for high-stakes binary security assessments.

Key insights

Semantic grounding in static and runtime evidence is crucial for reliable binary vulnerability detection.

Principles

• Ground context selection in memory flow evidence.
• Fuse dual IR/C representations selectively.
• Confirm feasibility with executable evidence.

Method

A semantic-driven Slicer extracts witness-backed flows; a dual-view Detector reasons over these flows; a multi-agent Validator confirms hypotheses via targeted execution.

In practice

• Lift binaries with RetDec.
• Use GPT-5.4 for detection.
• Validate with Radare2/Valgrind.

Topics

• Memory Corruption Vulnerabilities
• Binary Vulnerability Detection
• Large Language Models
• Agentic Frameworks
• Static Program Analysis
• Runtime Validation

Code references

• 0xdea/semgrep-rules
• AFLplusplus/AFLplusplus
• avast/retdec
• fkie-cad/cwe_checker
• KeenSecurityLab/BinAbsInspector

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Security Engineer, AI Scientist, Research Scientist

Related on AIssential

• Code-Augur: Agentic Vulnerability Detection via Specification Inference — Agentic vulnerability detection improves by making LLM assumptions explicit and continuously validating them with fuzzing.
• SoK: AI-Augmented Binary Reversing — AI-augmented binary reversing complements conventional methods, leveraging artifacts as an interface for learning-based semantic inference.
• A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle — LLM agent memory's inherent malleability necessitates lifecycle-wide security governance for mnemonic sovereignty.
• AI is getting scary good at finding hidden software bugs - even in decades-old code — AI excels at finding obscure bugs in legacy code but also expands the attack surface for unpatchable systems.
• Context-Fractured Decomposition Attacks on Tool-Using LLM Agents: Exploiting Artifact Provenance Gaps — Context-Fractured Decomposition (CFD) attacks exploit untracked artifact provenance in LLM agents, enabling delayed, multi-step jailbreaks.
• Building an Agentic Security Pipeline That Finds, Proves, and Patches Vulnerabilities — LLM-powered agentic pipelines can automate finding, proving, and patching vulnerabilities in complex codebases.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.