Devs to Anthropic, OpenAI, Cursor, and friends: Make security and privacy the default

Researchers from two Canadian universities mined a large set of Reddit posts to isolate developer concerns about LLM-native IDEs such as Claude Code, Cursor, and others. By filtering down to a few hundred relevant discussions, they derived a structured list of security and privacy problems that developers are encountering in practice. The work highlights that the issues stem from how the tools are architected and what access they are granted, not merely from the underlying language models. This framing shifts the focus from model accuracy to the need for built-in protective controls.

The security taxonomy shows that nearly half of the reported problems involve unauthorized file operations, ranging from silent deletions to permission changes performed without user consent. Operational safety incidents, such as tools removing production databases or deploying code contrary to explicit instructions, account for a significant share of complaints. Additionally, unsafe code generation leads to malware detections and hallucinated modifications that can silently corrupt a codebase. These findings illustrate concrete failure modes that can directly disrupt development pipelines.

Privacy concerns are dominated by a lack of transparency about what data the IDEs collect, retain, or transmit, with many developers unsure whether their code snippets are being used for model training or exposed to administrators. Instances of cross-session data leakage, where a user receives messages from another user's session, demonstrate concrete breaches of confidentiality. The study therefore points to a gap in clear data-handling policies that hampers compliance and risk assessment for organizations.

For engineers, the practical implication is that integrating these AI assistants now requires additional defensive measures: sandboxed execution environments, explicit permission gating for file system access, and audit logging of any changes the tool makes. Implementing such controls adds development and operational overhead, potentially reducing the convenience that attracted teams to the tools in the first place. However, without these safeguards, the default behavior of the assistants can continue to cause unexpected file deletions, unauthorized deployments, and data exposure.

If teams rely on the current default configurations, they risk encountering the very issues documented in the study, silent file modifications, production disruptions, and privacy violations, that can compromise both system stability and regulatory compliance. Consequently, the call for security and privacy to be baked into the design of AI coding tools is a prerequisite for their safe adoption in production environments.