I Made Two AIs Fight Each Other for 6 Months — Here's What They Learned

I Made Two AIs Fight Each Other for 6 Months — Here's What They Learned

In September 2025, I started an experiment: take two AI engines, put them in an arena with 51 layers of defense, and let them fight. No human intervention. Pure adversarial evolution.

One AI attacks. The other defends. They fight 24 hours a day. The loser evolves. The winner gets harder challenges. Repeat for six months.

What emerged was not what I expected.

The Setup

A genetic algorithm engine that breeds attack strategies. Each "generation" contains hundreds of candidate strategies. The fittest survive, crossover, mutate, and produce the next generation. Think biological evolution, but for cyber attacks.

The Sanglier doesn't know the rules. It doesn't have a playbook. It discovers strategies through pure trial and error across thousands of generations. Some strategies are elegant. Some are bizarre. Some are things no human would ever think to try.

A real-time defense engine that monitors, detects, and neutralizes attacks. It sits across 51 layers of infrastructure — each layer a different detection mechanism, a different defense strategy, a different trap.

The Spider learns from every attack. Every time the Boar finds a weakness, the Spider patches it and adapts. But it can't just patch — it has to predict. Because the Boar is evolving too.

51 defense layers. Think of it like a castle with 51 walls. The Boar needs to breach them. The Spider needs to hold them. Each layer has its own logic, its own vulnerabilities, its own strengths.

At the start, the Boar could breach about 15 layers. By month three, it could reach layer 30. By month five, it was consistently hitting layer 40.

The Spider, meanwhile, went from a 30% detection rate to 94%.

What the AIs Discovered

Discovery 1: Speed Beats Strength

The Boar's most successful attack strategies were not the most sophisticated ones — they were the fastest ones. Attacks that completed in milliseconds beat attacks that were technically superior but took seconds to execute.

The Spider confirmed this from the defense side: the single most important factor in successful defense was response time. A mediocre defense deployed in 100 milliseconds outperformed a perfect defense deployed in 5 seconds.

Discovery 2: Adaptation Trumps Optimization

The Boar could optimize a single attack strategy to near-perfection. But the Spider would learn to counter it. The most successful long-term Boar strategy was not optimization — it was rapid adaptation. Switching approaches every few generations, never letting the Spider calibrate.

The Spider learned the same lesson in reverse: the best defense was not building the strongest wall, but changing the wall structure frequently. Moving targets are harder to hit.

Discovery 3: Evidence Structure Matters More Than Evidence Quantity

This discovery directly led to the fintech breakthrough.

When the Spider successfully defended a layer, it had to "prove" the defense by generating a structured log. Early Spider versions dumped massive amounts of raw data — every sensor reading, every log entry, every timestamp.

Through evolution, the Spider learned that structured, concise evidence was far more effective than raw data dumps. It evolved a specific format:

1. 1. Summary of the attack (2 lines max)
2. 2. Key evidence (3-5 data points, most important first)
3. 3. Timeline (chronological, no gaps)
4. 4. Conclusion (1 line)

Sound familiar? This is exactly the format that wins chargeback disputes.

The Spider independently evolved the same evidence structure that top chargeback defense specialists use — because the underlying problem is the same: convince a reviewer, under time pressure, that your position is correct.

Discovery 4: The Attacker Reveals the Defender's Blind Spots

The Boar's failed attacks were as informative as its successful ones. By analyzing where attacks failed, the Spider could identify which defenses were strong and which were irrelevant.

Some defense layers that seemed important turned out to be useless — the Boar never even targeted them. Other layers that seemed minor were consistently probed.

Discovery 5: Co-Evolution Produces Superior Intelligence

After six months, both AIs were dramatically more capable than either would have been alone. The Boar was a better attacker than any human-designed attack system. The Spider was a better defender than any human-designed defense.

Neither could have reached this level without the other. Competition was the catalyst for intelligence.

This is the core insight of adversarial AI: intelligence doesn't emerge from data alone, or from architecture alone. It emerges from struggle.

The Pivot to Fintech

In February 2026, while analyzing the Spider's evidence generation patterns, something clicked.

The Spider had evolved a defense workflow that looked exactly like chargeback defense:

1. 1. Detect the threat in real-time (webhook notification)
2. 2. Classify the threat type (reason code)
3. 3. Gather forensic evidence (transaction data)
4. 4. Structure the evidence (formatted PDF)
5. 5. Submit the response (API call)
6. 6. Adapt based on outcome (learn from win/loss)

The Spider was already doing chargeback defense — it just didn't know it.

We connected the Spider's evolved intelligence to a Stripe payment processor. Fed it historical chargeback data. Let it evolve for a few weeks specifically on payment disputes.

The results were immediate: 60% win rate on disputes that merchants had been losing 100% of the time (because they never responded).

That system became ChargeShield.

What This Means for AI

We are living through a Cambrian explosion in AI applications. But most of the attention goes to large language models — chatbots, text generation, image creation.

The adversarial AI approach represents something fundamentally different: AI that learns by fighting, not by reading.

LLMs learn from human-generated text. Adversarial systems learn from competition. Both are powerful, but adversarial learning produces a different kind of intelligence — one that is strategic, adaptive, and battle-tested.

The most exciting AI applications of the next decade will likely come from combining these approaches: LLMs for understanding and communication, adversarial systems for strategy and adaptation.

Try It Yourself

The ChargeShield system — born from 6 months of adversarial AI combat — is available for Stripe merchants.

Connect your Stripe account, and the AI that learned to fight through 51 layers of cybersecurity defense will fight your chargebacks. Automatically. In 30 seconds.

No win, no fee. chargeshield.vmaxbadge.ch

Because the best defense is one that was forged in battle.

---