Practical Security for AI-Generated Code

Why It Matters

Without these safeguards, AI‑generated code can silently introduce vulnerabilities or be hijacked, exposing entire codebases and production environments to attack. Implementing minimal permissions, audit logs, and automated scans protects both security posture and developer velocity.

Key Takeaways

• Scope AI agent permissions to the minimum required resources.
• Implement detailed logging hooks to capture every agent action.
• Use automated code scanning tools for AI‑generated code.
• Integrate security checks across all development agents uniformly.
• Treat AI agents like interns: limit access, monitor, review continuously.

Summary

Milan Williams, product manager at Segrep, opened the session by warning that AI‑driven code generators are no longer limited to single‑line suggestions; they now produce thousands of lines of code and execute shell commands with elevated credentials. He framed the discussion around three practical safeguards that development teams can deploy today to keep AI agents from becoming security liabilities.

First, Williams urged teams to down‑scope access tokens, granting agents only the repositories and environments they truly need. He likened agents to new interns who should never receive unrestricted production access. Second, he emphasized the importance of an audit trail: simple logging hooks that capture every command and timestamp, as demonstrated with Cloud Code’s built‑in session logs and a four‑line script that records shell activity. Finally, he recommended automated code‑scanning solutions—language‑specific linters, Claude’s security bot, and Segrep’s free MCP server—that run on every line of AI‑generated code, ensuring vulnerabilities are caught before deployment.

Williams highlighted concrete examples: the Cloud Code project folder that stores session histories, a deterministic hook script distributed via MDM, and Segrep’s MCP server that integrates with major agents like Cursor, Cloud Code, and Windsurf. He noted that these tools require minimal setup yet provide critical visibility and protection, even when developers use different AI assistants.

The takeaway for enterprises is clear: treat AI agents with the same rigor as human contributors. By limiting permissions, logging actions, and scanning output, organizations can contain potential breaches, maintain compliance, and preserve the productivity gains promised by generative AI.

Original Description

Milan Williams (Semgrep) Lightning Talk at the Coding Agents Conference at the Computer History Museum, March 3rd, 2026.

Abstract //

AI is writing more code than ever, but Milan Williams warns most teams are basically handing agents the keys to production, so unless you lock down permissions, log everything, and scan outputs, you’re not moving faster—you’re just scaling security risks.

Bio //

Milan Williams is a Senior Product Manager at Semgrep, focused on developer-first security tools and bridging the gap between engineering and application security.