Research
Beware EviLLM:Enabling Vulnerability Injection via Large Language Models
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 24 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
EviLLM shows how account or browser compromise can undermine trusted AI coding workflows, reinforcing the need to protect those access paths and independently review generated code.
EviLLM is a demonstrated attack on AI code-generation pipelines that uses a compromised account or browser to deliberately introduce software vulnerabilities. Researchers implemented two variants covering 13 CWE classes; in a user study, 7 of 8 participants and 10 of 13 participants, respectively, did not notice the injected vulnerabilities.