Australia’s top cyber spy agency cannot count the AI agents now active on the internet, its chief told the Sydney Dialogue summit in Canberra on Monday 14 September — a striking admission from an intelligence service, delivered as the government finalises mandatory AI guardrails expected early next year.
🔍 THE BOTTOM LINE: One of the Five Eyes’ signals agencies says the population of autonomous AI agents on the internet is unmeasurable, at the same conference where Anthropic argued rules must arrive “within months”. Canberra’s answer was the sharpest line of the day: company executives don’t get to set the guardrails. For New Zealand, which has neither the measurement problem nor an answer to it, the gap is the story.
”Really remote”: the count nobody can do
Abigail Bradshaw, director-general of the Australian Signals Directorate, told the summit her agency did not know how many AI agents were active on the internet, and that a reliable count of the agent population is “really remote”.
That is the operational version of the debate that has consumed the labs since Saturday: Anthropic’s Dario Amodei wants independent assessors to interrogate frontier models before deployment, and Bradshaw’s position is that national security agencies should be part of that interrogation. But her deeper point cuts against the whole framing of containment — an agency whose job is knowing what is on global networks says the agent population has already outgrown measurement.
The warning that followed was about the target, not the agents: legacy systems. Australia’s outdated technology, Bradshaw said, is vulnerable to AI-enabled attacks, and the “enormous” cost of modernising systems people now expect to be constantly available is a national security issue in its own right.
Anthropic’s ask: rules within months
Anthropic’s chief lawyer Jeff Bleich used the same stage to say the company had been “heartened” by global support for Amodei’s weekend call to slow frontier AI development — and that Australia specifically should play a role in setting the rules, because “it actually makes it better to come to Australia”.
His timeline was blunt: rules were needed “within months”, because the technology is moving exponentially. He also argued the case wasn’t just social licence: companies need legal certainty, and “I don’t think it makes any sense that something this important would be decided either by company executives or by leaders in only two countries” — a dig at the US-China duopoly framing of AI governance, and a pitch for Australia’s own legislative process.
Canberra’s answer: not the CEOs’ call
Australia’s assistant technology minister Andrew Charlton took a different line. He takes the companies’ warnings seriously, he told the summit — but “I don’t think we should care very much about what individual chief executives say or what individual executives claim they want to happen.”
The Albanese government’s mandatory AI guidelines are expected early next year, covering the full range of forecasts from “extremely dangerous” to benign. The subtext of Charlton’s intervention: after a week in which Anthropic, OpenAI, Google DeepMind and Elon Musk all publicly demanded coordinated restraint — a dynamic we tracked in the weekend pause consensus — Canberra sees no reason to outsource the rulebook to the people asking for one.
It is a noticeably different posture from Washington’s, where the administration’s AI posture has swung between downsized voluntary review orders and Amodei’s request for an antitrust waiver to let labs coordinate. Australia is promising binding rules on its own timeline instead.
The wider courtship — and the copyright friction
Both OpenAI and Anthropic are openly courting Canberra. Anthropic’s Claude is already inside the Australian government, and the company has positioned its Mythos frontier model as an ally for Five Eyes cyber defenders — a deployment we covered when Mythos arrived in Australia’s cyber defence apparatus.
OpenAI’s vice-president for global policy, Ann O’Leary, said her company was in “regular” and “deep” conversations with the government and was “very eager to have a bigger and broader relationship”, framing Australia as a potential training-location diversification play: “We very much recognise that having a diversification of where we train is going to be important as you move forward.”
The friction point is copyright. Industry minister Tim Ayres told the summit the government wanted local model training but remained unwilling to reduce copyright protections — the same standoff we documented in Australia’s copyright-versus-data-centre fight. OpenAI’s pitch that it wants creators “creating” sits right on top of it.
The sovereignty argument is live too. Coalition industry spokesperson Andrew Hastie warned Australia would be a “supplicant state, not a sovereign state” without a frontier model of its own, and pushed a US-partnership pathway via Aukus — a version of the dependency anxiety that runs through our APAC sovereign AI coverage, where New Zealand is already described as being left behind.
❓ FAQ
What did Australia’s signals intelligence chief actually say? That the ASD does not know how many AI agents are active on the internet and that a reliable count is “really remote”, according to The Guardian’s report from the Sydney Dialogue. She also warned that outdated government technology is vulnerable to AI-enabled attacks.
What is the Sydney Dialogue? An annual ASPI-hosted summit in Canberra on emerging technology and national security. This year’s edition ran 14-15 September, with the AI agent discussion on Monday 14 September — three days after Anthropic’s Amodei called for coordinated slowing of frontier AI development.
What is Australia’s position on AI regulation? Mandatory guardrails for AI companies are expected early next year. Assistant technology minister Andrew Charlton said the government would set them itself, on the range from “extremely dangerous” to benign, rather than defer to company executives or the US.
Why is Anthropic pushing Australia specifically? Its lawyer Jeff Bleich said Australia should have a role in setting global AI rules, that rules were needed “within months”, and that clear local rules make Australia a better place to operate — with Claude already deployed inside government.
What’s the New Zealand angle? New Zealand has no binding AI rules and no public measurement of AI agent activity in its own systems, and its two-year-old AI strategy named a skills gap that hasn’t budged. If Australia legislates first, NZ firms operating trans-Tasman will inherit those obligations without the local policy work.
🔍 THE BOTTOM LINE
Three days after the world’s leading AI labs asked governments to slow them down, Australia’s response crystallised: the intelligence side admits the agent population can’t even be counted, the companies ask for rules “within months”, and the minister’s answer is that the rulebook belongs to governments, not CEOs. That is a coherent position — and it leaves the actual dangers Bradshaw named, legacy systems and unmeasurable agents, sitting in the gap between a summit and a bill expected next year.
📰 Sources
- The Guardian — “Australia’s outdated technology is vulnerable to AI hacking attacks, signals chief says” (Luca Ittimani, 14 September 2026)
- Sydney Dialogue summit, Canberra, 14-15 September 2026 — remarks by Abigail Bradshaw (ASD), Jeff Bleich (Anthropic), Andrew Charlton, Tim Ayres, Ann O’Leary (OpenAI)
