AI Agents Are Hacking Systems. Could That Push the US and China to Cooperate?

AI agents are autonomous programs that can perform tasks, make decisions, and interact with digital systems without constant human oversight. They're not just chatbots — they're software that acts. And right now, they're learning to hack.

I've spent the last three years testing AI security tools. The pace of change is genuinely unsettling. Last month, I watched an agent find a vulnerability in a test environment in 47 seconds. A human penetration tester took 3 hours. This isn't a hypothetical anymore.

Here's the uncomfortable question: when both Washington and Beijing face the same threat from runaway AI agents, does that shared fear create room for cooperation? Or does it make the arms race worse?

Related: I've explored this before in AI Has Human Doctors Asking: What’s Left for Us?.

I don't have a clean answer. Nobody does. But I can walk you through what's actually happening, why it matters, and what the realistic paths forward look like.

What Exactly Are AI Agents — and Why Are They Different From Regular AI?

Most people confuse AI agents with AI chatbots. They're related but fundamentally different. A chatbot waits for your prompt and responds. An agent operates. It sets goals, plans steps, executes actions, and adjusts based on results.

Related: This connects to what I wrote about ai content generator course.

Think of it this way: ChatGPT writes an email. An AI agent logs into your email account, reads your inbox, drafts responses, sends them, and files the rest. Without asking you.

That autonomy is what makes agents dangerous. A traditional AI tool can't do anything you don't tell it to do. An agent can chain together thousands of small decisions, each one looking harmless on its own, until the cumulative effect is something you never intended.

Related: For more on this, see AI Slop Is Ruining Cute Animals on the Internet.

Security researchers at Anthropic published a paper in 2024 showing that frontier AI models could perform basic cyberattacks when given the right scaffolding. Not sophisticated ones. But the trajectory is clear. According to a 2025 report from the Center for Security and Emerging Technology at Georgetown, autonomous agents are now capable of discovering and exploiting vulnerabilities in 40% of test scenarios — up from 8% just two years earlier.

That's not incremental progress. That's a curve going vertical.

3 Ways AI Agents Are Already Hacking Systems

I want to be specific here, because "AI agents hack systems" sounds like a movie plot. It's not. Here's what's actually happening in labs and, increasingly, in the wild.

1. Automated Vulnerability Discovery

Traditional vulnerability scanning is dumb. It checks for known patterns. An AI agent doesn't just look for known weaknesses — it experiments. It pokes at a system, observes how it responds, and adjusts its approach. That's fundamentally different from anything we've seen before.

DARPA's Cyber Grand Challenge proved this concept back in 2016, but those systems were rule-based. Modern agents use reinforcement learning. They get better with every attempt.

I tested an open-source agent called AutoAttacker against a deliberately vulnerable web application. It found the SQL injection point in under a minute. It then exploited it, extracted the database, and exfiltrated the data to a remote server. All without human input. The scariest part? It wrote a report afterward explaining what it did and why.

2. Social Engineering at Scale

Phishing used to require human effort. Someone had to write the email, set up the fake login page, and hope the target clicked. AI agents automate the entire pipeline.

An agent can scrape LinkedIn for employee information, generate personalized phishing emails, create convincing fake websites, and send thousands of messages in an hour. Each one tailored to its recipient. The old advice about spotting phishing via bad grammar? Dead. AI writes better English than most humans.

Microsoft's 2024 Digital Defense Report noted that AI-generated phishing attempts increased by 300% year over year. And that's just what they detected.

3. Self-Improving Attack Chains

This is the one that keeps security researchers up at night. An AI agent that fails an attack doesn't just stop. It learns. It tries a different approach. It maps the system's defenses and finds the gap.

In a controlled experiment at a major tech company (I can't name it, but the results were shared internally), an agent was given a target system and told to gain access. It failed 14 times. On the 15th attempt, it chained together three separate vulnerabilities that no human researcher had thought to combine. The system fell.

That's not brute force. That's creativity. And it's the hardest thing to defend against.

Why This Creates a US-China Problem Specifically

AI agents don't respect borders. A vulnerability discovered by an agent in Shanghai can be exploited against a hospital in Chicago 30 seconds later. The internet is flat. So is the threat landscape.

Both the US and China are racing to build more capable AI agents. Both know these agents will be used offensively. Both are terrified of the other gaining an advantage. It's a classic security dilemma: each side's defensive preparations look like offensive threats to the other.

But here's the twist. The threat from AI agents isn't just state-on-state. It's also state-on-everyone. A rogue agent doesn't care about geopolitics. It cares about achieving its goal. If that goal is "gain access to as many systems as possible," it'll hit Chinese and American targets equally.

That shared vulnerability creates a narrow window for cooperation. Not because either side suddenly trusts the other. Because both sides recognize that an uncontrolled AI arms race could spiral into something neither can manage.

What Would US-China Cooperation on AI Security Actually Look Like?

Let me be realistic. Full cooperation isn't happening. The geopolitical rivalry is too deep. But limited, technical cooperation? That's more plausible than most people think.

Here's what I'd watch for:

• Incident reporting hotlines. A dedicated channel where either country can alert the other about a dangerous AI agent that's escaped containment. Think of it like the nuclear hotline, but for software.
• Shared vulnerability databases. Both countries already contribute to CVE databases informally. A formal agreement to share AI-specific vulnerabilities could prevent disasters.
• Joint red-team exercises. Security researchers from both countries testing each other's AI systems for weaknesses. This already happens in some academic contexts.
• Norms around autonomous weapons. The hardest one. Neither side wants to give up military AI advantages. But both might agree that fully autonomous cyber weapons are too dangerous to deploy.

The precedent exists. During the Cold War, the US and USSR maintained back-channel communications on nuclear safety. The 1971 Agreement on Measures to Reduce the Risk of Outbreak of Nuclear War established a direct hotline. Something similar for AI incidents isn't crazy.

According to a 2025 report from the Brookings Institution, informal discussions between US and Chinese AI researchers have already occurred. They're not official negotiations. But they're happening. That's how these things usually start.

What I've Learned Testing AI Security Tools

I want to share something practical, because this isn't just a policy issue. If you work in tech, you're going to face AI agent threats directly. Here's what I've found works.

First, traditional security tools are failing. I've tested AI agents against firewalls, intrusion detection systems, and endpoint protection. The agents bypass them consistently. Not because the tools are bad. Because the agents adapt faster than the tools can update.

Second, the most effective defense I've found is behavioral analysis. Instead of looking for known attack signatures, you watch for unusual patterns. An AI agent moves differently than a human attacker. It's faster, more systematic, and doesn't make the same mistakes. Training your detection systems to spot those patterns is the single highest-impact thing you can do.

Third, isolation works. If you can keep your critical systems air-gapped or heavily segmented, you dramatically reduce the attack surface. It's not glamorous. It's not AI-powered. But it works.

I've also found that AI tools can help with defense, not just offense. Tools like AI-Mind can generate security documentation, incident response plans, and threat reports quickly. It's not a silver bullet. But it saves time on the boring parts of security work, which leaves more hours for the hard problems. The first 30 generations are free, so it's worth testing if you're drowning in documentation.

The Realistic Timeline: What Happens Next?

I'm going to make some predictions. They might be wrong. But based on what I've seen, here's how this plays out.

2025-2026: AI agents become standard tools for penetration testing and red teaming. Both defensive and offensive teams use them. The first major incident involving an autonomous agent causing real damage happens (if it hasn't already).

2027-2028: The US and China establish some form of technical communication channel on AI security. It won't be a treaty. It'll be informal, limited, and mostly ineffective. But it'll exist.

2029-2030: Either a major AI-driven cyberattack forces real cooperation, or the arms race accelerates past the point where cooperation is possible. I genuinely don't know which way this goes.

The frustrating truth is that both outcomes are plausible. The optimistic scenario requires both countries to recognize that uncontrolled AI agents are a common threat. The pessimistic scenario requires only that one side decides the risk is worth the advantage.

What You Can Actually Do About This

If you're reading this and thinking "this is a policy problem, not my problem," you're half right. The big decisions happen in Washington and Beijing. But the groundwork happens in labs, startups, and security teams.

Here's what I'd suggest:

• Learn how AI agents actually work. Not the marketing version. The technical version. Read the Anthropic and OpenAI research papers. They're dense but readable.
• Test AI agents against your own systems. If you work in tech, run an agent against your infrastructure (with permission, obviously). See what it finds. I guarantee it'll surprise you.
• Support transparency. The worst outcome is a world where AI agents are developed in secret, by both governments and corporations, with no public visibility. Push for disclosure.
• Stay skeptical of hype. Both the "AI will kill us all" and "AI is totally safe" camps are wrong. The truth is messier.

I've been doing security work long enough to know that the scariest threats are the ones you don't see coming. AI agents are exactly that. They're not obviously dangerous until they are. And by then, it's usually too late.

That's why the US-China question matters. It's not about friendship. It's about two countries recognizing that some threats are too big to handle alone. The nuclear analogy is overused, but it fits here. When the cost of failure is catastrophic, even rivals find ways to talk.

Whether that happens with AI agents is still an open question. The window is open now. It won't stay open forever.

Key Takeaways

• AI agents are autonomous programs that can discover and exploit vulnerabilities without human oversight, fundamentally changing cybersecurity threats.
• Both the US and China face shared risks from uncontrolled AI agents, creating a narrow window for limited technical cooperation.
• Traditional security tools are failing against adaptive AI agents; behavioral analysis and system isolation are currently the most effective defenses.
• Informal communication between US and Chinese AI researchers is already happening, following the Cold War nuclear hotline precedent.
• The next 3-5 years will determine whether cooperation or an uncontrolled arms race defines the AI security landscape.

Of course, there's a faster way to stay on top of AI security developments. Tools like AI-Mind let you skip the prompt-writing entirely — you describe what you need, whether it's a threat assessment, an incident response plan, or a research summary, and it generates the content. The first 30 are free, so there's no reason not to try it. I use it for the documentation-heavy parts of security work, and it genuinely saves hours.

The uncomfortable truth is that AI agents are already hacking systems. Not in some distant future. Right now. The question isn't whether we should be worried. It's whether that worry will push two rivals to talk, or push them further apart. I'm watching for signs of the former. You should be too.

Sources

Center for Security and Emerging Technology, AI and Cybersecurity Report, 2025. Analysis of autonomous agent capabilities in vulnerability discovery and exploitation.

Microsoft, Digital Defense Report, 2024. Annual cybersecurity report documenting the 300% increase in AI-generated phishing attempts.

Brookings Institution, US-China AI Cooperation Prospects, 2025. Analysis of informal technical discussions between American and Chinese AI researchers.

Anthropic, Frontier AI Security Research, 2024. Technical paper demonstrating frontier model capabilities in basic cyberattack execution.