Key Takeaways

  • Nikesh Arora points out that models trained on code hallucinate 30% of the time when finding bugs, but attackers still win because they only need one working exploit.
  • Arvind Krishna reports that coding agents write roughly 50% of IBM's internal code, outperforming human engineers by running continuous evaluations before pull requests merge.
  • Mathematical verification of software across broad inputs remains computationally infeasible, leaving automated testing agents as the only viable line of defense.
  • Arora argues software vendors must pool vulnerability discoveries through trusted parties to distribute patches globally rather than forcing companies to patch open-source flaws individually.
  • Krishna counters asymmetric attack economics by applying Krishna's Three Strategic Principles of Cyber Defense Warfare.

Krishna's Three Strategic Principles of Cyber Defense Warfare

Models trained to write software understand code structure so well that they excel at breaking it. Nikesh Arora observed this dynamic directly: “If you teach a model what good code looks like, it also can figure out what bad code looks like. And if you let it loose against your code, it'll tell you what is bad.” Even with a high error rate, offensive agents hold a structural edge. “What's fascinating is when it did that, even then it was hallucinating 30% of the time,” Arora noted. “The reason it's perfect for attackers, they just need one vulnerability. They don't need 10 vulnerabilities.”

Defenders cannot rely on formal mathematical proofs to stay safe. Arvind Krishna explained that complete verification across broad inputs remains “computationally infeasible right now.” Instead, defense requires changing the economics. “Technique two in warfare what has been the most effective is asymmetry,” Krishna explained. “Often an attacker is cheaper than the defender. So the defender runs out of money and at some point then concedes. Or can you make defense equally cheap, in which case you can basically have symmetric warfare and then you have a much better chance of success.”

Krishna organizes cyber defense into three strategic principles:

Technique 1: Powerful Defensive Alliances

Form broad defensive alliances that share and pool vulnerability discoveries, building a unified defense network where trusted partners distribute patches globally so that the alliance becomes more powerful than any individual attacker.

Technique 2: Asymmetric and Low-Cost Defense

Neutralize the attacker's cost advantage. Attackers typically operate far cheaper than defenders; to prevent defenders from depleting capital and conceding, defensive measures must be made cheap and efficient enough to achieve symmetric warfare.

Technique 3: Absolute Technological Superiority

Secure technological superiority over the adversary by leveraging advanced AI agents to identify vulnerabilities, auto-patch systems, and establish boundaries that make penetration computationally and financially prohibitive.

When This Works (and When It Doesn't)

This framework applies when enterprise software and digital infrastructure face automated, machine-speed attacks from well-funded adversaries or hostile states. When opponents script agents to probe endpoints continuously, manual audits fail. Defense must run autonomously at near-zero marginal cost per check, or the defender goes broke.

It breaks down inside small codebases built on custom, proprietary architecture. Defensive alliances rely on shared libraries and common open-source dependencies; isolated stacks receive no pooled patches. At the same time, if your testing pipeline lacks rigorous evaluation benchmarks, defensive agents will flood developers with false alarms, matching the 30% hallucination rate Arora identified.

What to Do With This

If you run a startup shipping code daily, apply Krishna's three techniques to your repository this week:

First, implement Technique 1 by connecting your repository to automated upstream dependency feeds. Stop auditing open-source libraries by hand. Let trusted vendor networks flag and patch supply-chain flaws automatically.

Second, apply Technique 2 by slashing verification costs. Add an automated security agent to your CI/CD runner to review pull requests before human review. Krishna noted that IBM generates “probably half the code that is written inside our place” with agents because “the agents can also do eval, basically make sure that everything they can find is actually fixed before it ever goes out.”

Third, enforce Technique 3 by scheduling an adversarial agent to attack your staging environment nightly. Have it hunt for API authorization bugs and unescaped inputs, then automatically file a pull request with the fix.