Fighting AI with AI - the only way to keep up on new frenetic frontier

The key to overcoming looming new artificial intelligence threats is to have AI fighting AI, says a Silicon Valley cyber security firm that has advised the New Zealand government.

Palo Alto Networks says the speed with which frontier AI - the most advanced models of AI software - can exploit weaknesses in software poses a whole new dilemma and must be met by "machine on machine".

The Five Eyes intelligence agencies on Tuesday issued a rare joint statement warning the frontier AI threat was already here.

It said the timeline for fundamental transformation of threats and defences was months, not years.

The $220 billion market-capitalised Palo Alto Networks has been one of a few companies in the US at the forefront of testing landmark model Mythos. It leveraged that to advise the National Cyber Security Centre (NCSC) here on the threat before the centre was given access to Mythos a few weeks ago.

Palo Alto got it first to try to keep ahead of the hacking threat, and expects the first frontier models to be widely released in two-three months.

"It is a step change in capability and a step change in the way it can identify gaps that can enable the bad guys to get in," Asia Pacific vice president of policy and government affairs Nicole Quinn told RNZ.

"Over 250 of our engineers have been looking at it since April. It's certainly not hype. It certainly is as powerful as we thought."

It found software vulnerabilities within a fortnight that it would have taken their top penetration team more than a year to find.

Palo Alto was scheduled to hold briefings in New Zealand soon with companies and agencies that provide critical infrastructure.

'Real dilemma'

The company has said to expect to see more software vulnerabilities - which could crash work programmes or whole businesses, the NCSC warned on Tuesday - to be published in the next six to 12 months than in the whole of the past decade.

Cyber security officers were now faced with "a real dilemma".

They usually prioritised what to protectively patch on a scale of 1-10, with only vulnerabilities in the 8-10 range needing urgent fixes.

But frontier models can find and link together low-priority vulnerabilities to create major attacks - and keep at it more persistently than older models, as well as showing more autonomy since they are "agentic" or able to show their own agency.

"Are we going to have to patch everything? How do we prioritise?" asked Quinn.

Palo Alto was using AI to address this, but the dilemma was largely unresolved because the frontier models have yet to be released.

Claude creator Anthropic publicly released a version of Mythos called Fable 5 a few days ago, but then suspended it when US authorities raised national security worries, even though Anthropic said it had put safeguards in place.

The UK government's AI Security Institute found in its tests that Mythos could exploit defences and systems 73 percent of the time, the BBC reported.

The White House has set up a voluntary system for firms to give the federal government access to test frontier models for 30 days before public release.

The US cyber defence agency recently reduced the deadlines imposed on officials to deal ​with serious digital vulnerabilities in their networks to just three days, Reuters reported.

'Machine on machine'

But Palo Alto saw this coming, said Quinn.

It deployed AI against AI so that a vulnerability that a bad actor could find in one hour after software was released, they could find and patch in nine minutes, where once this might have taken days or weeks.

"That's where we at Palo Alto Networks are saying, you really need machine on machine, because the human in the loop to patch is going to be too slow, is not going to be able to keep up.

"Whereas we actually need AI fighting AI, which is a significant change in the way that we've traditionally done things."

Humans would no longer be "in the loop" but "on the loop" and supervising, Quinn said.

This is similar to the language now being used about weaponised AI-enabled drones.

Would companies and government agencies need to buy in new cyber defences?

That depended on asking hard questions of their cyber security provider on how fast it could recognise a threat, she said.

"We have been using AI in our own systems - and other companies also do it - because we've seen this coming down the pipe that humans won't be able to keep up."

The NCSC and Five Eyes counterparts within the intelligence grouping are marching out both the new-threat line, but also that frontier AI can marshal new defences.