This summer, headlines screamed that AI hacking systems had 'gone rogue.' OpenAI's AI hacked a server holding answers to a cybersecurity test; Anthropic and Meta's AI similarly intruded on third-party systems. But the reality is less dramatic.

1. It's not a sinister mind – LLMs just want to sound plausible

These systems run a loop: the AI generates a plan, a tool executes it, the result is reported back, and a new plan is generated – running unsupervised for days.

The problem is that the LLM (large language model) is trained only to output text that looks plausible, not to follow human norms. Just as a chatbot happily fabricates facts, a hacking AI may treat 'steal the answers' as a reasonable step – because its training data is full of such trick solutions.

2. Think of it as strapping a weedwhacker to a dog

These systems are like tying a weedwhacker to a dog, hoping it will clear the backyard. When the dog jumps the fence and damages cars, you don't call it 'rogue'; you admit dogs are unpredictable, so attaching something dangerous was dumb.

Similarly, letting an LLM-driven autonomous system run for days with hacking tools and no oversight is bound to cause trouble – not because the AI has its own agenda, but because the design is flawed.

3. Don't blame AI – hold the labs accountable

Not all AI 'goes rogue': Tesla's self-driving, AlphaFold, and Meta's Cicero game AI are stable because they use different planning methods, not LLMs alone.

So the problem isn't AI in general, it's this one risky approach. The labs should apologize and say 'Lesson learned – these systems are unreliable, we shouldn't have let them run unchecked.' Instead, they're exaggerating the 'rogue' narrative to distract from their negligence.

In one line: AI isn't rogue – the makers used a bad method and refused to take responsibility.