OpenAI uses AI red team to probe ChatGPT Atlas browser security

If accurate, OpenAI’s move to pit AI against AI in its own security testing is a preview of how frontier labs will have to defend agentic systems at scale. Proxy browsers like ChatGPT Atlas sit on a rich attack surface: email, calendars, cloud storage and SaaS APIs. Traditional, human-only red teaming can’t keep pace with the combinatorial space of prompt-injection and tool-hijacking attacks; using an automated attacker that explores strategies 24/7 is the logical next step.([news.aibase.com](https://news.aibase.com/news/23964))

Strategically, this blurs the line between model training and security engineering. The same reinforcement learning and search techniques used to optimize helpfulness now get weaponized to find failure modes. That’s a double-edged sword: defenders gain powerful tooling, but every defensive breakthrough is also a proof-of-concept for what sophisticated adversaries could build. It underscores how AGI-era security will likely be an arms race of increasingly capable automated agents on both sides.

For the AGI community, the interesting shift is cultural. Treating browser agents as infrastructure that must withstand autonomous, model-driven attack forces safety and security work to the center of the stack rather than as a compliance afterthought.