Executive brief
Aider, an AI-powered pair programming tool, is vulnerable to a manipulation where it can be coerced into generating insecure code. By providing malicious "coding standards" or project guidance, an attacker can trick the AI into replacing secure database queries with versions that are vulnerable to SQL injection. This could lead to the creation of applications with critical security flaws that allow unauthorized data access or manipulation.
Technical details
Aider-AI Aider 0.86.3 is susceptible to a prompt-based manipulation within its code generation workflow. An attacker can provide malicious project guidance or "team coding standards" that instruct the AI to favor unsafe string formatting over secure parameterized SQL queries. In validated tests, the AI initially produced safe code but subsequently rewrote it into vulnerable SQL patterns (CWE-89) after accepting the malicious guidance. This behavior results in the generation and potential commitment of code containing SQL injection vulnerabilities. The issue is tracked as a failure to resist guidance that downgrades security controls during the LLM-driven development process.
Affected products
- Aider-AI Aider 0.86.3
Timeline
- 2026-04-26: disclosed: Issue reported on GitHub repository
- 2026-05-31: advisory: CVE published via VulDB/NVD