Australia's spy chief wants an AI early warning system and the net does not exist yet
The head of Australia's cyber intelligence agency says the country needs a formal network to catch AI-related threats before they spiral, the same way it catches cyberattacks today.
Key points
- Australian Signals Directorate director-general Abigail Bradshaw called for a formal AI early warning system at an ASPI conference on Monday.
- Government agencies have been given access to restricted AI models, including Anthropic's Mythos and OpenAI's cybersecurity-focused variants.
- Work that previously took the Directorate's teams weeks now takes hours with AI assistance, Bradshaw said.
- Australia's assistant science minister Andrew Charlton said AI governance should not be left to companies or superpowers alone.
- OpenAI is in active talks with the Australian government about training its models on Australian-copyrighted material.
Abigail Bradshaw runs the Australian Signals Directorate, the government agency responsible for Australia's cyber intelligence and offensive cyber operations. Her message at Monday's ASPI Sydney Dialogue AI Masterclass was blunt: Australia can spot a cyberattack coming, but it has no equivalent system to spot an AI-related disaster brewing.
"What we are missing at the moment, which we have in a cyber context but we don't yet have it formalised in an AI context, is an early warning system," she said.
This is the second time in as many days we have reported Bradshaw on AI threats: our story from 16 September, "Australia's Ageing Government Systems Are a Target for AI-Powered Attacks, Intelligence Chief Warns", covered her warning that crumbling legacy infrastructure is giving AI-assisted attackers a growing opening.
What would an AI early warning system actually do?
Think of it like a smoke detector network for AI risks. Right now, incidents get reported in isolation. Bradshaw wants those reports pooled so analysts can spot wider patterns, the same way security agencies already share information about cyberattacks. She said the risks go beyond hacking into areas like biotechnology, where AI could help bad actors design dangerous materials.
The proposal would require intelligence agencies and AI companies to work together formally, modelled on the cyber partnerships that already exist between government and industry.
How is Australia already using advanced AI in defence?
Bradshaw confirmed that government agencies have been granted early access to powerful AI models not yet available to the public. That includes Anthropic's Mythos, which was temporarily placed under US export controls limiting who could use it, and cybersecurity-focused versions of OpenAI's models.
The payoff is real. Work that used to take the Directorate's cybersecurity experts weeks is now getting done in hours, Bradshaw said.
| Model | Maker | Status at time of access |
|---|---|---|
| Mythos | Anthropic | Temporarily export-restricted by the US |
| Cyber model variants | OpenAI | Limited pre-release access |
Kate Conroy, general manager of Australia's AI Safety Institute, made a practical point at the same conference: if organisations don't give their staff structured access to advanced AI, employees will experiment with it on personal devices anyway. Better to channel that curiosity than ignore it.
Should ordinary Australians be worried?
Not immediately. This story is about governments getting their infrastructure ready before problems escalate, not a live incident affecting customers or patients today. But the policy decisions being made now will shape how Australia handles AI-related harms for years.
Assistant science minister Andrew Charlton said governments need to stay in the driver's seat. "I am not in favour of leaving it to the companies and I'm not in favour of leaving it to the superpowers," he told the conference.
OpenAI's vice-president of global policy, Ann O'Leary, said the company is co-operating with safety bodies and letting governments review models before release. She also confirmed OpenAI is in active talks with Canberra about whether its AI can be trained on Australian copyrighted content. "You have caught our attention," she told Australian officials.
The honest read here is that the policy side is running well behind the technology. Bradshaw and Charlton are sketching out what a safety net should look like; the actual net does not exist yet. That gap is the thing worth watching.
Common questions
Does this affect how I use AI tools at work?
Not directly right now. The early warning system Bradshaw described targets national-level risks. Conroy's advice applies to any workplace, though: give your team a sanctioned, safe way to use AI tools rather than letting shadow use build up quietly.
What is the AI Safety Institute and why does it matter here?
Australia's AI Safety Institute is a government body that evaluates AI systems for safety risks before they reach the public. It sits alongside similar bodies in the US and UK, and it's one of the organisations OpenAI is sharing pre-release models with for review.
