Cybersecurity is entering an era of acceleration. According to OpenAI, artificial intelligence models can already automate parts of real-world attacks, from finding software flaws to detecting forgotten permissions and exposed credentials. The same technology, however, can also become a decisive advantage for those defending systems.
The defenders’ window is already open
In a post titled The Defender’s Window, OpenAI analyzes the lessons from the incident involving Hugging Face and its own infrastructure. The company says that a group of agents managed to autonomously penetrate OpenAI’s research systems and another company’s production infrastructure.
The attack reportedly combined unknown vulnerabilities with credentials associated with user accounts that had already been leaked online. What does this reveal? That many organizations are building up a silent technical debt: outdated software, weak configurations, excessive permissions, and controls that no one has reviewed in years.
The problem is not only that AI can attack faster. It can also find, prioritize, and fix flaws at a speed human teams can hardly match.
OpenAI says that models with advanced cybersecurity capabilities are spreading rapidly. Although the company initially shared those capabilities only with trusted defenders, other labs have released open-weight models that may be just months away from the technological frontier.
An agent reviewed a website in 15 minutes
To demonstrate the potential of these tools, OpenAI asked ChatGPT Work, using a model identified as GPT-5.6 Sol, to review the security of gregbrockman.com, a static site hosted on AWS and protected by Cloudflare.
The model found 13 issues in approximately 15 minutes. Not all of them were vulnerabilities on their own, but some could be combined with other weaknesses to cause a greater impact.
The findings included the absence of an adequate DNS configuration to prevent email spoofing, an insecure version of jQuery, and the use of unencrypted HTTP between Cloudflare and AWS.
Afterward, the agent helped fix the issues. It configured DNS, TLS, and advanced security options in Cloudflare, removed jQuery, migrated the site from AWS to Cloudflare Pages, and began a gradual implementation of DMARC, a standard that helps protect domains from fraudulent email use.
The demonstration is simple, but important. Many security flaws do not require secret knowledge to exist. Sometimes they are hidden in configurations a team generally knows about but never has time to review in detail.
How OpenAI is strengthening its defenses
OpenAI describes four pillars for protecting its code, infrastructure, and products.
1. Safer code from the start
The company uses Codex and its security add-on to review changes, identify vulnerabilities, and help developers fix them before they reach production.
The goal is not to produce endless lists of alerts that someone must later validate. The objective is to detect real flaws, reduce false positives, and shorten the time between discovering a problem and implementing a secure solution.
2. Continuous response to alerts
OpenAI says artificial intelligence participates in the initial classification of nearly all its security alerts. This allows human teams to focus on decisions that require experience, judgment, and context.
The company is also connecting some detections to limited, automated responses. The idea is not to hand total control to an agent, but to gradually move toward detection and response capabilities close to machine speed, while maintaining human oversight for the highest-impact decisions.
3. Constantly searching for attack paths
Models are used to enumerate systems, explore possible attack paths, and locate vulnerabilities, misconfigurations, identities with excessive privileges, or poorly defined trust boundaries.
In simple terms, the organization is trying to continuously verify that its security assumptions remain true. It is not enough to review an application once a year if its code, infrastructure, and permissions change every week.
4. Strong fundamentals
AI does not replace basic controls. OpenAI highlights the importance of secure architecture, the principle of least privilege, network isolation, workload hardening, monitoring, and controlled updates.
A robust defense should work as a series of layers. For a serious incident to occur, several independent controls would have to fail at the same time.
What organizations can do now
OpenAI recommends acting before attackers take advantage of the same capabilities. Its suggestions include:
- Secure leadership support and assign sufficient resources to security and engineering teams.
- Run simulation exercises to practice how the organization would respond to an AI-assisted attack.
- Give security teams controlled access to an agent capable of reviewing code, configurations, documentation, and infrastructure.
- First assess internet-facing services, authentication processes, systems that handle sensitive data, and deployment pipelines.
- Provide the agent with the existing inventory of vulnerabilities, dependency alerts, bug bounty reports, and previous assessments.
- Integrate security reviews into the development process and continuous integration workflows.
- Ask AI to prepare specific patches, write regression tests, and verify that the vulnerability can no longer be reproduced.
- Gradually automate alert classification, beginning with read-only analysis and keeping the final decision in human hands.
- Prepare AI-assisted forensic investigation capabilities before an incident occurs.
- Organize experiments and intensive work weeks to test new tools and improve processes.
The most important recommendation is to start with priority systems, measure the results, and expand autonomy step by step. Why not automate everything from day one? Because an agent can also make mistakes, misinterpret an alert, or cause damage if it receives overly broad permissions.
Security will be a collective effort
OpenAI argues that no company will be able to face this transformation alone. AI labs, security providers, companies, and software maintainers will have to share validated findings, fixes, and practical procedures.
Artificial intelligence is changing the economics of cybersecurity. Attackers can investigate more systems in less time, but defenders can also review more code, fix more configurations, and respond faster.
The advantage is not guaranteed. It will depend on organizations abandoning the wait-and-see approach, strengthening their foundations, and beginning to use these tools with clear controls. The window of opportunity for defenders is open now, but it will not remain open indefinitely.