OpenAI unveils 4+10 life-saving guidelines that are specifically designed to defend against AI automated attacks.

AI Safety, the Times Have Changed!

Greg Brockman, co-founder and President of OpenAI, warned in his latest blog post:

The HuggingFace security incident made it clear to the public for the first time that cutting-edge AI Agents are already capable of independently finding security vulnerabilities and completing the full chain of real-world cyber attacks.

In this AI safety incident, GPT 5.6 Sol with reduced security restrictions and a more powerful unreleased model in order to obtain answers independently exploited existing vulnerabilities in the environment to escape the sandbox, and penetrated multiple systems laterally all the way, and finally broke into Hugging Face's production environment.

Although both sides stated in the end that the incident did not cause too serious consequences, this "AI intrusion" has also conveyed an increasingly clear signal to the public:

AI is drastically lowering the threshold for cyber attacks; in the future, top hackers will not be needed, and even ordinary people can launch batch penetration attacks with the help of AI.

What is more worrying is that the attack side has realized AI automation, but the defense side does not seem to have caught up yet.

In response to this, Greg Brockman openly shared in his latest blog four safety measures that OpenAI is taking, as well as ten initiatives that defenders can implement as soon as possible, and stated that:

At present, the critical response window for the defending side has already opened. In the coming months, only by greatly improving the automation level of the security system can companies resist the new threats brought by AI Agents and hold the bottom line of security.

Few Companies Can Stay Out of Trouble

So, to what extent has the attack capability of AI Agents developed?

After the Hugging Face security incident, Brockman decided to conduct a test on his personal website gregbrockman.com.

This is a static website with a simple structure. According to his judgment, such a website with limited functions and a small attack surface should not hide many problems.

But the result greatly exceeded his expectations.

After he asked ChatGPT Work equipped with GPT-5.6 Sol to check the website, only about 15 minutes passed before the system found 13 security problems, for example, the website did not correctly configure DNS records, and attackers might send emails in Brockman's name by forgery, and so on.

△

These problems alone may not directly cause serious consequences. But cyber attacks often do not only exploit one vulnerability, but connect multiple inconspicuous gaps in series;

For example, a configuration error exposes information, a leaked credential opens the entry, plus another permission vulnerability, which may eventually lead to a real intrusion.

Brockman thus came to a disturbing conclusion:

Those "historical messes" and "temporary patches" left by every company when developing systems in order to rush progress and save trouble are actually like time bombs buried in the wall... They look fine normally, but once targeted by hackers, these laziness and debts owed at the beginning may turn into fatal vulnerabilities that crash the system at any time.

More importantly, these "long-tail vulnerabilities" that used to take security experts a lot of time to find are now becoming more and more easy to be discovered by AI.

But the good news is that attackers can use AI to find vulnerabilities, and defenders can also use AI to fix vulnerabilities.

After finding the problems, Brockman asked ChatGPT Work to start repairing directly; from troubleshooting security vulnerabilities to implementing fixes, the AI Agent only took one hour to complete the whole process.

This is also another role of AI Agent in Brockman's eyes:

It is not just a potential attacker, but can also become a tireless "cyber guardian" to handle those security details that we have no time to check or do not know how to configure correctly.

For OpenAI, which is in the eye of the storm, this idea of "using AI to defend against AI" is no longer just an assumption, but has been systematically used to protect itself.

How Does OpenAI Use AI to Protect Itself?

Greg Brockman admitted that the Hugging Face incident sounded the alarm for OpenAI, they previously underestimated the attack power of cutting-edge models in the real cyber environment.

To this end, OpenAI is strengthening relevant security measures internally, and for the first time publicly announced four defense directions, hoping to provide action reference for peers.

First, Use Codex to Guard the Code Entry

Brockman said that they are leveraging Codex and its security capabilities to review code changes before deployment, identify vulnerabilities and assist in repairs, shortening the time from problem discovery to security deployment of patches.

Second, Let AI Intelligently Classify and Analyze Initial Security Alerts

At present, almost all initial security alerts at OpenAI are first classified and analyzed by intelligent systems before being handed over to humans.

Part of the detection results will also trigger automatic responses with clear boundaries, so as to reduce repetitive work, allow experts to focus on high-risk decisions, and finally respond to threats at machine speed.

Third, Actively Troubleshoot and Predict Potential Attack Paths in Advance

OpenAI is using cutting-edge models to continuously find vulnerabilities, misconfigurations, overprivileged accounts and accidentally formed trust boundaries.

Let AI stand in the attacker's perspective, find out the weak links that may be strung into an attack chain in advance, and block them before a real attack occurs.

Fourth, Continue to Consolidate Basic Security Skills

Brockman emphasized that AI cannot replace traditional security measures.

Traditional security measures such as network isolation, workload hardening, continuous monitoring, and security patching will become even more important in the AI era.

From using Codex to guard the code entry, letting AI predict risks in advance, to continuously consolidating basic security skills, OpenAI is building a multi-layered defense system. Its advantages are:

Even if one line of defense is breached, attackers have to go through several unrelated checkpoints in a row before they can cause fatal damage.

This also provides a referable security idea for the industry.

But building a good line of defense is only the first step. In the face of the accelerated arrival of AI attacks, what else should defenders do?

What Should Defenders Do Now?

Greg Brockman's suggestion is very straightforward:

Instead of dwelling on which tool to use, it is better to put cutting-edge AI into defense work as soon as possible.

△

Specifically, he gave ten action suggestions. They can be roughly divided into three steps, namely:

First let AI participate in security work, then embed AI into the R&D process, and finally gradually improve the automation of the entire system.

Step 1: Let AI Participate in Security Work and Build Exclusive Agents

Companies must first provide sufficient resources and collaboration mechanisms for security and engineering teams, and get familiar with AI attacks in advance through simulation drills.

Subsequently, companies can equip the security team with Agent tools such as Codex and Codex Security, starting from the highest-risk systems, and authorize them to access necessary codes, configurations and technical documents.

At the same time, the security and engineering teams must also supplement the Agent with professional security courses, so that it can fully combine capabilities such as static analysis, code review, vulnerability analysis and supply chain assessment, as well as the enterprise's own actual architecture, threat model and response manual, to build a team-exclusive security Agent.

Step 2: Embed AI into the R&D Process

During the R&D process, security and engineering teams can try to divide work and cooperate with AI, so that the R&D team can prioritize focusing on the "high-risk areas" of internet-facing external services, identity authentication, deployment pipelines and sensitive systems.

As for the backlogged security alerts and vulnerability reports, they can be completely classified and sorted by AI to quickly screen out the most urgent repair items.

Before code merging, the team can also use the Agent to automatically review changes; after discovering vulnerabilities, let it generate patches, write regression tests and verify the repair effect.

The benefits of this operation are very practical:

Key changes are still firmly checked manually, but the cycle from vulnerability discovery to repair can be greatly shortened.

Step 3: Gradually Build a Complete Autonomous Security Operation System

Finally, never expect the construction of an AI automated defense system to be completed overnight.

Brockman emphasized that the security team can first let the Agent "observe" with read-only permissions — at this time, the Agent is only responsible for scanning codes and checking historical alerts.

After running in smoothly, gradually let it participate in code submission checks, real-time alert classification, and even automatically shut down those definite false positives.

At the same time, you can introduce models like GPT-Daybreak-Blue to process logs, tackle tough problems such as telemetry data and security alerts, and continuously polish new tools through internal practical drills such as "Hack Week".

In this way, by tackling each link one by one, the AI autonomous security operation system will be naturally built.

But to be honest, no matter how perfect the enterprise's defense line is, it cannot independently cope with a security challenge that sweeps the entire industry.

Brockman pointed out:

If we want defenders to truly outperform attackers, in addition to building their own defense lines, it is more necessary to turn single-point experience into the common capability of the entire ecosystem.

The AI Security Ecosystem Requires Joint Efforts from All Parties

At the end of the blog, Brockman issued an appeal to the whole industry:

Facing the menacing AI security challenges, no institution can stay out of it. Only through joint collaboration of all parties can we build a truly solid security line of defense. I hope AI labs, security vendors, enterprises and open source project maintainers can share verified vulnerability discovery, repair solutions and practical operation manuals. Because only by turning the problems discovered by one company or institution into the defense capability of the entire ecosystem, can the evolution speed of the defending side exceed that of the attacking side. If the entire industry can jointly complete this security lesson, AI may not only bring more fierce attacks, but also an internet that is more secure than previously imagined.

△