ETH Zurich researchers find coding agents copy security flaws already in a codebase into new code

arXiv · Research preprint

Read the source (opens in a new tab)
What happened
Researchers at ETH Zurich, including Martin Vechev, posted a preprint on November 8, 2026 testing whether vulnerable existing code leads AI coding agents to write similar vulnerabilities.
Key points
  • On 20 paired C/C++ tasks from real projects, all five agents tested did worse when nearby code was insecure; the drop averaged 35.4 percentage points and ranged from 25.0 to 49.0.
  • GPT-5.6 Sol's rate of correct and secure code fell from 91.0% with secure surrounding code to 42.0% with insecure code. On 96 backend tasks, insecure neighboring apps cut correct and secure code by 3.4 percentage points versus having no neighbors.
  • Telling agents to be careful about security weakened the effect and telling them to follow existing code style strengthened it, but neither stopped it.
  • Asked to explain, agents cited another code site in 76.5% of cases while recognizing the flaw they introduced in 69.9%; the authors warn code bases "can increasingly degrade in security."
Evidence
Research preprint, not peer reviewed, testing five frontier models in coding-agent harnesses on controlled tasks with paired secure and insecure code. Claude Opus 5.5 and GPT-6 Astra ran only the C/C++ tasks.

See it in the latest record