All problems
AI Safety
Adversarial Robustness of Code Generation Models
Code: ANX-RP-005
Study how LLM-based code assistants can be manipulated to generate vulnerable or malicious code, and develop defenses.
Why it matters
As developers increasingly rely on AI code assistants, adversarial prompts that generate insecure code pose a systemic risk to software supply chains.
Research gap
No systematic study exists on adversarial attacks targeting code generation models in real-world development workflows.
Research questions
- 01 What prompt modifications cause code models to generate SQL injection vulnerabilities?
- 02 Can we detect adversarial inputs to code models at inference time?
- 03 How do fine-tuning defenses compare to input sanitization approaches?
Suggested technologies
Required skills
Team and supervisor
2-3 researchers, supervised by Muhammad Qasim.
Suitable for: CS, AI, SE.
Expected deliverables
- Adversarial benchmark for code generation models
- Defense mechanism with provable robustness guarantees
- Security audit framework for AI-assisted codebases
Publication
Adversarial Attacks on LLM Code Assistants: A Systematic Study
Technical Report · 2026-09-01
Work on this problem
Apply to the lab and this problem becomes your supervised research project.