OpenAI's Chief Scientist: Using Stronger AI to Defend Against AI Threats Is the Strongest Argument for Accelerating Training

OpenAI Chief Scientist Jakub Pachocki recently published an article titled An Alien Mind on the company's official blog, laying out key thoughts on finding a balance between AI safety and development. His perspective reveals a central paradox in the AI field today: we need more powerful AI to defend against the very risks that AI itself poses.

Pachocki succeeded co-founder Ilya Sutskever as OpenAI's Chief Scientist in 2024. Sutskever had gradually stepped back from the company following the November 2023 board upheaval, eventually founding SSI (Safe Superintelligence Inc.), a company focused on safe superintelligence research. Pachocki had previously led OpenAI's post-training team for an extended period and played a key role in the development of GPT-4 and the o1 models. His decision to publish this article at such a sensitive moment—amid ongoing external scrutiny of OpenAI's safety commitments—serves as both a statement of technical philosophy and a piece of strategic communication.

Why AI Defense Is the Core Driver for Accelerating Model Training

Pachocki states explicitly that the single strongest argument for continuing to rapidly train more intelligent models is building defense systems to counter the dangers posed by other AIs. This perspective breaks away from traditional AI safety discussion frameworks, elevating defensive capability to a strategic priority.

In his view, the future AI safety ecosystem must rely on powerful and aligned AI systems to achieve three key objectives:

• Protecting critical infrastructure: Using AI to monitor and harden key infrastructure in real time, defending against potential AI-driven attacks
• Countering rogue agents in real time: Deploying defense systems capable of rapidly identifying and neutralizing malicious AI agents
• Inventing entirely new safeguards: Leveraging AI's creativity to design safety mechanisms beyond human imagination

The term "alignment" here refers to the technical and theoretical framework for ensuring that an AI system's goals, behaviors, and values remain consistent with human intentions. The core challenge is that as AI capabilities grow, even small deviations in objectives can be amplified into catastrophic outcomes. Current mainstream alignment approaches include Reinforcement Learning from Human Feedback (RLHF), Constitutional AI, and Scalable Oversight. OpenAI had previously established a dedicated Superalignment team and planned to devote 20% of its compute to this research direction, but the successive departures of key team members exposed deep internal tensions between safety and the pace of commercialization.

On the technical front, AI threats to critical infrastructure are far from hypothetical. The cybersecurity field has already observed real-world cases of AI being used to automate vulnerability discovery, generate phishing content, and even assist in writing malicious code. Both the U.S. National Security Agency (NSA) and the Cybersecurity and Infrastructure Security Agency (CISA) have published dedicated reports on AI threats. The "AI vs. AI" defense concept operates across multiple technical layers: using large language models for code auditing and vulnerability prediction, deploying AI-driven intrusion detection systems for real-time anomaly identification, and using AI to simulate attack scenarios for Red Teaming exercises. However, defenders perpetually face an asymmetric disadvantage—attackers only need to find a single vulnerability, while defenders must protect every possible attack surface. This asymmetry is the technical root of the defensive urgency Pachocki describes.

Pachocki emphasizes that this will become OpenAI's top deployment priority. In other words, OpenAI isn't just training more powerful models—it's building a comprehensive AI safety ecosystem.

Balancing Speed and Caution: OpenAI's Perspective

Despite emphasizing the necessity of rapid development, Pachocki simultaneously issues a sober warning: uncertainty and the need for defense cannot become excuses for recklessness. He points out that once you truly grasp the gravity of AI development, the idea of "charging ahead at all costs" becomes absurd.

This stance reflects deep internal reflection at OpenAI on the pace of AI development. Since ChatGPT's launch in late 2022, the global AI field has entered an unprecedented period of competitive acceleration. OpenAI, Google DeepMind, Anthropic, Meta AI, and multiple Chinese companies are all racing to release more powerful foundation models. GPT-4, released in March 2023, demonstrated multimodal understanding capabilities; the o1 series of models introduced "chain-of-thought reasoning" mechanisms, achieving significant breakthroughs in complex tasks like mathematics and programming. Meanwhile, Anthropic launched its Claude series and pioneered the concept of a "Responsible Scaling Policy," while Google pushed forward in multimodal capabilities with its Gemini series. This race extends beyond the technical dimension into talent acquisition, compute stockpiling, and geopolitical competition. Against this backdrop, safety researchers widely worry that commercial pressures could lead to compressed safety testing and weakened safety guardrails.

Pachocki's position attempts to chart a middle course between two extremes—neither freezing in place out of fear, nor blindly chasing speed while ignoring risk.

The Logical Dilemma of Defensive Accelerationism

Interestingly, the logic Pachocki proposes contains a potential self-reinforcing loop: accelerating the development of stronger AI to defend against AI threats, while that stronger AI may itself introduce new threats requiring even stronger defenses. This "defensive accelerationism" could become a perpetual justification for pushing AI capabilities ever higher.

Defensive accelerationism is an emerging school of thought in AI policy discussions, situated between "effective accelerationism" (e/acc) and "AI pause" advocacy. Effective accelerationism advocates for unrestricted technological advancement, holding that technological progress is inherently the greatest good; the AI pause position (as called for in the open letter initiated by the Future of Life Institute) advocates halting large-scale AI training until safety research catches up. Defensive accelerationism attempts a middle ground: acknowledging the reality of AI risks while arguing that the best way to address them is not to slow down, but to ensure that "good AI" stays ahead of "bad AI." This reasoning shares similarities with deterrence theory in the military domain, but critics note that the equilibrium logic of nuclear deterrence may not apply to AI, since AI capabilities spread and can be replicated far faster than nuclear weapons.

From a tech ethics perspective, this framing raises even deeper questions:

• Who gets to define "alignment"? Different organizations may hold fundamentally different standards for AI alignment. OpenAI, Anthropic, and DeepMind each have distinct safety research paradigms, and academia and industry don't always share the same priorities. Without unified standards, each company may equate its own technical approach with "safety" itself.
• Whose defense system can be trusted? Defensive capability is itself a form of power concentration. The organization that controls the most powerful defensive AI effectively gains the power to define the AI safety narrative—which, in the absence of effective external oversight, could create new power imbalances.
• How do we ensure defense systems aren't misused? With AI capabilities advancing rapidly, can oversight mechanisms keep pace? Any sufficiently powerful defense system also possesses offensive capabilities in technical terms, and this dual-use nature makes a purely "defensive" positioning difficult to sustain in practice.

Far-Reaching Implications for the Industry and AI Regulation

Pachocki's statements mark an important expression of OpenAI's official position and provide a noteworthy reference framework for the entire industry. It's foreseeable that AI companies will increasingly emphasize the safety and defense capabilities built into their new models going forward.

This perspective may also influence the direction of AI regulatory policy. If "using AI to defend against AI" becomes the dominant narrative, regulators may need to find a new balance between encouraging safety research and controlling risk proliferation. The global AI regulatory landscape is currently evolving rapidly: the EU's AI Act began phased implementation in 2024 with a risk-tiered regulatory framework; the U.S. leans toward a combination of executive orders and industry self-regulation; and China has also issued multiple regulations targeting generative AI and algorithmic recommendations. Across these different regulatory paths, how to assess the boundary between "defensive AI R&D" and "offensive AI capability accumulation" will become a thorny challenge that legislators in every country must confront.

For AI practitioners, Pachocki's perspective sends a clear signal: technological progress is not just about expanding capabilities—it's about shouldering greater responsibility. The pursuit of more powerful models must be accompanied by the concurrent development of corresponding safety mechanisms and ethical frameworks. As he puts it, understanding the severity of the risks is the prerequisite for avoiding reckless behavior.

Key Takeaways