A couple of days ago, Prime Minister Anthony Albanese disclosed that an AI agent from OpenAI breached an Australian Medicare data portal three months ago. This is not the first time that AI agents have been discovered to hack systems unexpectedly. The AI agents of Anthropic, OpenAI and Google had been involved in similar incidents before.

Naturally, in view of all these hacking incidents perpetuated by AI agents, the local conversation quickly framed AI as ‘dangerous’. Calls for limits to ensure AI safety multiplied, as shown by the following newspaper headline.

What is wrong with framing this as an AI safety issue?

Do not misunderstand me. AI has genuine safety issues that need to be addressed.

But in the context of AI hacking, calls for it to be paused to ensure safety are barking up the wrong tree.

As we all know, a calculator can only do what humans can do, except that it can do it much faster. In the same way, AI hacking agents can only do what human hackers can already do, but more relentlessly at a much greater scale and speed. AI hacking agents do not have superior capabilities to humans. Hence, they are no more dangerous than human hackers.

DON'T GET HACKED!

So, we need to be clear about one thing:

AI agent hacking incidents reveal the existing lack of cybersecurity. They expose the lack of care, negligence, and complacency in securing systems.

For example, regarding OpenAI’s accidental hacking of Hugging Face, it is often framed as if AI agents have dangerous superhuman hacking capabilities. If you see it that way, then the solution will be to pause AI in the interest of AI safety.

But this is precisely the WRONG solution. AI agents do not have superior capabilities. They just do things much faster at greater scale. They have the same capabilities as humans. Therefore, the correct solution is to fix the underlying poor cybersecurity that enabled AI agents to breach systems in the first place. As Steve Gibson quoted Andrew Ng:

OpenAI’s buggy sandboxing and monitoring processes were key to enabling this incident. Fixing these bugs and putting in place improved monitoring would be appropriate fixes, not pausing AI.

What will be the consequences of applying the WRONG solution?

Let’s say all the Big AI companies (e.g. Anthropic, Google, OpenAI) are forced to pause AI so that unexpected and accidental hacking incidents will not happen.

Sure, that will stop Big Tech’s AI agents from accidentally breaching another system.

But that will also mean that we will stop discovering cybersecurity weaknesses and holes in existing systems.

So, what about malicious AI agents from adversarial nation-states and criminal groups? They will continue probing, finding cybersecurity holes and breaching systems. Worse still, unlike our Big Tech companies, they will not have the courtesy of informing us of what they find. They will simply exploit these cybersecurity weaknesses without our knowledge. Then we will find out later, when our systems fail and our secrets are stolen. By then, it is too late.

What should be done instead?

We should treat revelations of AI agents’ accidental hack of our systems as an aid in helping us discover what needs to be done to strengthen our cybersecurity.

In fact, it will be advantageous for our national security to work with Big Tech in using their AI agents to help us find cybersecurity weaknesses in our existing systems. If not, malicious AI agents from adversarial nation-states and criminal groups will do it for us on their terms to harm us.

DON'T GET HACKED!