AI Coding Tools Must Embed Security to Handle Autonomous Agent Risks
Security cannot remain a final checkpoint in software delivery if AI agents are generating hundreds of code changes daily. Ox Security's field CTO, Boaz Barzel, told attendees at Infosecurity Europe on June 4 that traditional application security-built around monthly pen testing cycles-no longer fits the speed of agentic development.
"Security isn't a stage in the pipeline; it's a property of the act of creation itself," Barzel said. "We have to shift into the agent."
Four New Attack Surfaces
AI agents create four distinct entry points that existing security tools don't address:
- Input: Instructions entering the agent from developers, upstream agents, or threat actors
- Tools: MCP servers, models, skills, and external SaaS connections-authorized or shadow-that could be weaponized to steal data, inject instructions, or move laterally
- Execution: Autonomous agents running without visibility, enforcement, or accountability
- Output: Vulnerable or destructive code generated at machine speed without human review-path traversal, injection attacks, backdoors, data exfiltration logic
The risk window is narrowing. Powerful frontier models could reduce time-to-exploit from discovery to attack in minutes. Combined with the volume of code AI tools generate, the scale of exposure has fundamentally changed.
Security as System Behavior
Barzel outlined an "auto-pentest loop" where security agents work alongside coding agents. Every commit gets tested. Every fix gets validated autonomously. The system reasons about what changed, what's exposed, and what risk it introduced.
"Security stops being a department. It becomes a behavior of the system," he said.
This approach aims for specific outcomes: mean time to resolve vulnerabilities dropping from weeks to hours, 100% autonomous coverage of merged changes, and most issues fixed without human intervention.
Real-World Vulnerabilities Emerging
New risks surface regularly. In May 2026, researchers discovered a critical vulnerability in the Cline Kanban server that could allow threat actors to silently hijack AI coding tools.
Development teams using AI coding tools and generative code systems should assess whether security checks run continuously during development, not just before deployment.
Your membership also unlocks:
