Newest papers

SPLASH: Contrastive Audit Pressure for Detecting and Localizing Security Hallucinations in Generated Code

Mirza Masfiqur Rahman, Ajwad Akil, Imtiaz Karim, Ashish Kundu, Ramana Rao Kompella, Elisa Bertino · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26001 · cited by 0 works (OpenAlex) · open-access PDF

Indonesian

Kode yang dihasilkan oleh model bahasa besar sering melewati setiap tes fungsional sementara tetap tidak aman. Halusinasi keamanan ini umum, ditulis dengan keyakinan tinggi, dan diulang di berbagai sampel, sehingga tidak ada ketidakpastian generator sendiri atau analis statis yang dapat diandalkan menangkap mereka. Kami memperkenalkan SPLASH, yang dipandu oleh langkah baru yang disebut tekanan audit. Hal ini kontras dengan probabilitas log yang dipaksa oleh guru untuk program yang dihasilkan di bawah pemeriksaan keamanan dan pemeriksaan gaya. Dikombinasikan dengan pernyataan dan program, dan didasarkan pada model probabilistik dari editing yang bersaing, tekanan audit memungkinkan SPLASH untuk memilih program yang aman di antara program kandidat model dan menemukan pernyataan yang rentan, tanpa pelatihan model, aturan khusus CWE, atau akses internal ke model generator. Pada 63.684 program yang dihasilkan oleh 12 model untuk

Machine translation (Meta M2M100); the English original follows.

Abstract

Code generated by large language models often passes every functional test while remaining insecure. These security hallucinations are common, written with high confidence, and repeated across samples, so neither the generator's own uncertainty nor static analyzers reliably catch them. We introduce SPLASH, which is guided by a new measure called audit pressure. It contrasts teacher-forced log-probabilities for the generated program under a security review and a style review. Aggregated to statements and programs, and grounded in a probabilistic model of competing edits, audit pressure allows SPLASH to choose the secure program among a model's candidate programs and locate the vulnerable statement, with no model training, CWE-specific rules, or internal access to the generating model. On 63,684 programs generated by 12 models for 899 tasks, SPLASH picks secure code in 74% of cases, places the vulnerable statement first in 58% of insecure programs, and chooses better than a frontier LLM judge.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Code (set theory), Audit, Probabilistic logic, Computer security, Measure (data warehouse)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange