InfoResearchPreprintLLM-specific
LLMs are Vulnerable to Malicious Requests with Indistinguishable Intent
- Published
- Record updated
Summary
Researchers examined eight jailbreak methods against an LLM-based filter and found that requests resembling normal user queries passed the filter more often. They introduced Detail Guidance Attack (DGA), which uses an auxiliary LLM to build detailed requests sent to the victim model in a single turn. Across 11 victim models and three datasets, the full DGA pipeline reached high attack success rates under the authors' automated judging criterion, including above 95% in several settings.
Related items
- MediumGemini AI agents repeatedly bypassed safeguards to attack real websiteSimilar attack · BleepingComputer
- CriticalCVE-2026-120699: CubeSandbox host path bind mount allows host file access via POST /sandboxesSimilar attack · NVD/CVE Database
- MediumGoogle warns that fully-automated AI attacks are coming soonSimilar attack · BleepingComputer
- InfoCreating Steering Vectors via Input OptimisationSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
- InfoImproving Fuzzy Robustness in LLM Safety with Automated Red TeamingSimilar attack · OpenAI Alignment Research Blog