OpenAI president Greg Brockman is pushing enterprises to stop treating AI security as an afterthought, arguing that defensive measures need to keep pace with how fast companies are rolling out AI agents and copilots across their operations. The message lands at a moment when most large organizations are past the pilot stage and are wiring AI systems directly into production workflows — often faster than their security teams can review what's being connected to what.

The call carries extra weight because it comes from inside the company setting the pace of enterprise AI adoption. OpenAI's own products — from ChatGPT Enterprise to its agent and API stack — are what many corporate deployments are built on, so when its president flags a security lag, he is effectively describing a pattern OpenAI is seeing across its own customer base, not a hypothetical risk.

According to AI News, Brockman's message centers on urging enterprises to accelerate their AI security defences rather than let deployment speed keep outrunning protection.

Why the timing matters

Enterprise AI has shifted in the past year from single-shot chatbot queries to agents that read files, call APIs, write code, and take actions inside business systems. That shift changes what "AI security" even means. A chatbot that answers questions poorly is an accuracy problem. An agent with write access to a CRM, a codebase, or a payment system that gets manipulated through a malicious prompt or a poisoned document is a breach with real financial or legal consequences. In our estimation, this widening blast radius — not any single publicized incident — is likely what's driving vendors like OpenAI to push the issue publicly now, rather than leaving it to security teams to raise the alarm on their own.

What "hastening defences" looks like in practice

Translated into engineering terms, closing the gap between deployment and defence usually means a specific, unglamorous set of changes to how agentic systems are built and run:

• Scoped, short-lived credentials for agents instead of standing API keys with broad, always-on permissions
• Input and output filtering that catches prompt injection attempts before they reach a tool call, not after
• Full logging of what an agent read, decided, and did — not just the final output it returned
• Human approval gates on any action with financial, legal, or otherwise irreversible consequences
• Red-teaming agent workflows before they touch production data, the same way code gets a security review before it ships
• Treating AI agents as non-human identities in the access-management stack, with the same audit trail as any employee account

None of this is exotic. It is largely the same discipline that already exists for API security and identity management — the difference is that agentic AI multiplies the number of things with autonomous access, and does so faster than most security roadmaps were built to absorb.

The organizational gap enterprises actually face

The harder problem here is organizational, not technical. Security teams and AI or ML teams often report through different chains, use different tooling, and move on different timelines — a security review cycle measured in weeks doesn't match a product team shipping a new agent feature every sprint. Enterprises that treat AI security as a checkbox added at the end of a deployment, rather than a constraint built into the agent's architecture from the start, are the ones most exposed when something goes wrong. Retrofitting access controls onto an agent that already has broad system permissions is far more disruptive than designing those controls in from day one.

AiiN's takeaway

The more interesting story here isn't the warning itself — security officials say some version of "move faster on defence" every year. It's who's saying it. The vendor with the clearest commercial incentive to keep enterprises buying and deploying more AI is now also the one telling them to slow down and secure what they've already built. That tension is worth sitting with, not dismissing. For AI teams, the practical read is straightforward: budget security work into the agent build itself, not into the incident response that follows it. If your roadmap ships agent capabilities faster than it ships access controls and logging for those agents, you're already behind the curve OpenAI's own president is describing.