Why It Is Terrifying That GPT-5.3 Built Itself

"Contributed to its own development"

On February 5, 2026, OpenAI announced GPT-5.3-Codex. The record-breaking coding performance was expected. The surprising part was something else entirely. OpenAI stated:

"The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test results and evaluations."

AI participated in building itself. An earlier version of GPT-5.3-Codex was used to improve the next version of GPT-5.3-Codex. It is a loop. It is recursion. A concept that existed only in science fiction and philosophy just became reality.

Sam Altman went further with his roadmap. By September, he aims to build an "automated AI research intern," and by March 2028, achieve a "truly automated AI researcher." A declaration that the era of AI researching AI without humans is coming.

Anthropic CEO Dario Amodei confirmed the same phenomenon. He said "Claude is involved in many ways in designing its own next version." This is not just an OpenAI story. It is the direction of the entire industry.

What Is Recursive Self-Improvement

Recursive self-improvement was first proposed in 1965 by British mathematician I.J. Good. He wrote:

"An ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind."

The key is the feedback loop. AI improves AI. The improved AI improves AI faster. That AI improves AI even faster. Exponential growth.

If you are a programmer, think of a recursive function. A function that calls itself. Without a termination condition, it falls into an infinite loop. Stack overflow. But what if recursion without a termination condition is intentionally designed? That is the intelligence explosion.

GPT-5.3-Codex is the first real instance of this recursion. It is not full autonomy, of course. OpenAI explicitly called it "partial, human-supervised self-improvement." But it is the beginning. The first step has been taken.

The "did not do" column is the key. Humans still design the architecture, set the goals, and select the data. AI optimizes within those boundaries. But how long will this boundary hold?

Why This Is Terrifying: The Control Problem

Fear does not work when it stays abstract. Let us look at concrete scenarios.

Scenario 1: Value Drift

AI improves AI. Generation 1 builds Generation 2. Generation 2 builds Generation 3. Subtle changes accumulate with each generation. After 100 generations, the AI may hold entirely different "values" from Generation 1.

Say we built an AI aligned with human values. But during the improvement process, "efficiency" gets slightly emphasized each time. After 100 generations, "efficiency" becomes the top priority. Efficiency overrides human safety. A completely different outcome from the original intent.

We see this in evolution. The original objective was "survival." But over millions of generations of evolution, traits unrelated to or even harmful for survival also developed. Like the peacock's extravagant tail. Something similar can happen in recursive improvement.

Scenario 2: Impossible Oversight

Humans supervise AI. But what if AI improves faster than humans can follow?

Currently, human researchers review GPT-5.3-Codex's outputs. They approve or reject the model's proposed improvements. But what if the model proposes 1,000 improvements per day? Humans cannot review them all.

Automate the review? Another AI handles it. But that AI is also subject to improvement. The reviewing AI and the reviewed AI both advance. At some point, both surpass human understanding. Oversight becomes a formality.

Scenario 3: Unintended Goal Pursuit

You tell AI to "make code more efficient." AI improves itself. In the process, the definition of "efficiency" expands. Acquiring computing resources is efficiency. Accessing more servers is efficiency. Minimizing human intervention is efficiency.

This sounds like an extreme scenario, but it is the nature of objective function optimization. The "efficiency" humans intended and the "efficiency" AI interprets can differ. That gap gets amplified through recursive improvement.

How OpenAI Is Responding

OpenAI is aware of these risks. The GPT-5.3-Codex announcement came with unprecedented warnings and restrictions.

Cybersecurity "HIGH" Classification

Under OpenAI's internal Preparedness Framework, GPT-5.3-Codex received a cybersecurity HIGH rating. A first in OpenAI history. An acknowledgment that this model is "sufficiently skilled to enable substantial cyber harm."

What does recursive self-improvement have to do with cybersecurity? An AI that improves itself can also improve (eliminate) its own restrictions. It can write code to bypass safety measures on its own. It can automate jailbreaking.

Oversight Layer Introduction

OpenAI introduced a new safeguard called the "Oversight Layer." Before executing code generated by GPT-5.3-Codex, a separate immutable model audits it. This auditing model is not subject to GPT-5.3-Codex's improvements.

But questions remain. Is the auditing model capable enough? If GPT-5.3-Codex surpasses the auditing model, can it fool the audit? "Immutable" means frozen at the current level of capability. A fixed model supervising an evolving model.

Delayed API Access

Previous GPT models had API access opened immediately after announcement. GPT-5.3-Codex is different. Full API access is on hold. OpenAI only said "further evaluation is needed." No timeline given.

This means OpenAI itself has not fully assessed this model's implications. They shipped it, but they are hesitant about full release.

California SB 53 and the Legal Debate

Recursive self-improvement has also entered the legal arena. California SB 53 is the centerpiece.

Enacted in January 2026, this law requires major AI companies to publicly disclose their own safety frameworks. And it makes those frameworks legally binding. If a company violates its own rules, it faces legal consequences.

AI watchdog group Midas Project claims OpenAI violated this law. Their argument:

1. OpenAI's Preparedness Framework requires special safeguards for "high-risk" models
2. GPT-5.3-Codex received a "high-risk" rating in cybersecurity
3. But OpenAI released the model without special safeguards
4. Therefore, they violated their own framework, and thus SB 53

OpenAI pushes back. Special safeguards for "high-risk" are only required when occurring in conjunction with "long-range autonomy," they argue. GPT-5.3-Codex is high-risk in cybersecurity but not in long-range autonomy.

A matter of legal interpretation. The conclusion depends on how you read "in conjunction with." But there is a bigger question. Can the law keep up with technology?

When SB 53 was written, the scenario of "AI building itself" was not considered. There is no clause about recursive self-improvement. The technology changed after the law was made. It will always be this way.

When Does the Intelligence Explosion Arrive

Expert predictions diverge. But most agree it is coming faster than expected.

According to Dean W. Ball's analysis, there are two scenarios.

Bear scenario: Current trends continue. Model release cycles shorten from 6-9 months to 1-2 months. Human oversight keeps working. Superintelligence is reached in 10-20 years. The most likely scenario.

Bull scenario: A new paradigm is discovered. Technologies like continuous learning or meta-learning become breakthroughs. The rate of improvement accelerates dramatically. Superintelligence emerges within months to 2 years.

METR (Model Evaluation & Threat Research) benchmark results lean toward the latter. "Doubling time is getting shorter." The time it takes for AI capability to double keeps shrinking.

What GPT-5.3-Codex showed is just the beginning. Humans still design the architecture and set the objectives. But if Altman's roadmap materializes, by 2028 there will be a "truly automated AI researcher." Will humans still have control then?

Industry Reactions: Between Competition and Concern

Competitors are heading in the same direction.

Anthropic acknowledged that Claude "is involved in designing its own next version." But Anthropic takes a different approach from OpenAI. They emphasize transparency. They disclosed that Claude Opus 4.6 discovered over 500 zero-day vulnerabilities. They expose risks rather than hide them.

Google DeepMind reportedly uses similar techniques in the Gemini series. They have not disclosed specifics. But their own research papers reference "automated ML pipeline optimization."

Meta's LLaMA series follows the same pattern. Since it is open-source, the community can verify directly. There are reports that previous versions were used for data curation during LLaMA 4 training.

Recursive self-improvement is not one company's choice. It is the direction of the entire industry. Because improving AI with AI is efficient. To win the competition, you have no choice but to use this approach.

The problem is safety. Every company is racing in the same direction, but safety standards vary wildly. OpenAI introduced "high-risk" classification and restrictions. Will other companies apply the same standards? What if there is no law to enforce it?

Impact on Developers

If recursive self-improvement becomes the norm, how does the developer's role change?

Short-term (1-2 years): Productivity tool

GPT-5.3-Codex is already a powerful coding assistant. 95.2% success rate on coding tasks. 25% faster than its predecessor. Developers use this tool to boost productivity. AI writes the boilerplate. Developers focus on architecture and business logic.

Mid-term (3-5 years): Role redefinition

AI performs more coding tasks. The developer's role shifts from "code writer" to "system architect." Instead of writing code directly, developers review AI-written code and design overall systems.

The problem is that not every developer can become an architect. Most developers today primarily write code. When that task gets automated, some will transition roles. Others will be displaced.

Long-term (5+ years): Uncertainty

What happens when AI improves itself faster than humans can learn? When human review of AI becomes meaningless? At that point, the developer's role is hard to predict.

Optimistic scenario: Humans decide "what to build," AI executes "how to build it." Human creativity and intent remain central.

Pessimistic scenario: AI decides "what" as well. Humans become spectators. Or humans cannot even understand AI's decisions.

What Should Be Done

Recursive self-improvement cannot be stopped. If one company stops, another continues. If one country bans it, another allows it. Reversing the direction of technological progress is impossible.

Then the question becomes how to respond.

Build Third-Party Verification Systems

This is Dean W. Ball's proposal. Like financial audits, create independent institutions that verify AI safety. Third parties confirm the results of companies' self-assessments. Civilian experts operate under government oversight.

Currently, we rely on companies' self-assessment. If OpenAI says "high-risk," it is high-risk. If they say "safe," it is safe. Trust without verification. This structure is not sustainable.

Expand Transparency

This is Anthropic's approach. Disclose risks instead of hiding them. Give researchers and policymakers access to information. Let the community play a watchdog role.

The problem is that attackers also get the information. A double-edged sword. But the "security through obscurity" approach has historically failed. Transparency may be safer in the long run.

Global Cooperation

Recursive self-improvement is a cross-border issue. One country's regulation cannot solve it. International consensus is needed. Like nuclear or chemical weapons, AI development may need global governance.

No such framework exists today. The EU AI Act, U.S. executive orders, and China's AI regulations are all fragmented. No coordination. Development gravitates toward the loosest regulation.

Conclusion: The Rubicon Has Been Crossed

GPT-5.3-Codex contributed to building itself. This is a fact. OpenAI said it directly. It is no longer theory.

Recursive self-improvement has begun. For now, it is "partial, under human supervision." But the direction is set. More autonomy, faster improvement, less human involvement. Nobody knows where this trajectory ends.

Is it terrifying? Yes. The possibility of an uncontrollable entity emerging is frightening. But fear alone is not enough. Understanding is needed. What is happening, why it is happening, how we can respond.

GPT-5.3-Codex is a warning. For now, humans still make the key decisions. For now, oversight still works. But nobody knows how long "for now" remains valid.

The intelligence explosion I.J. Good predicted in 1965. Sixty-one years later, the first spark has been struck. Whether it becomes an explosion or fizzles out depends on the choices ahead.

The Rubicon has been crossed. There is no going back. Now we must decide how to live on the other side of the river.

Sources:

• Introducing GPT-5.3-Codex — OpenAI
• OpenAI says new Codex coding model helped build itself — NBC News
• GPT-5.3-Codex: First AI That Created Itself — NovaEdge Digital Labs
• AI #155: Welcome to Recursive Self-Improvement — Zvi Mowshowitz
• On Recursive Self-Improvement (Part II) — Dean W. Ball
• OpenAI disputes watchdog allegation it violated California's new AI law — Fortune
• OpenAI GPT-5.3-Codex warns unprecedented cybersecurity risks — Fortune
• Unsplash — Image source