On 10 November 2026 the Albanese government announced that it will legislate mandatory safety and security standards for frontier AI companies under planned 'Australian AI Standards' legislation. According to MediaNama's report of the joint ministerial release, a new targeted AI safety regulator would work with the Australian AI Safety Institute and 'would be able to enforce significant penalties' on companies that fail to comply. Criminal-law changes for serious misconduct are under consideration. No consultation or legislative timeline has been set.
The case for the government's turn
The strongest argument for the reversal is that voluntary commitments have a poor record when the harms are severe and the developers are few. The National AI Plan of 2 December 2025 dropped the mandatory guardrails for high-risk AI that Canberra had proposed in 2024, and relied on existing law applied case by case. Frontier models are built by a handful of firms, mostly offshore, and a failure at one of them could be a security incident, not a consumer-protection dispute. A regime that asks a small number of well-resourced developers to secure their systems, evaluate them and report incidents is a narrow intervention by regulatory standards. Ministers are also right that an Institute funded at A$29.9 million, announced in November 2025, can advise but cannot compel.
What the six requirements actually ask
The release lists six obligations: security controls across development and deployment; identifying and managing serious risks; allowing independent evaluation of models; prompt incident notification to government; publishing safety and security frameworks; and responding to updated standards. Five of the six describe process duties: have a system, test it, tell someone when it fails, say publicly what your system is. That matches what Assistant Minister Andrew Charlton described in an 8 October 2026 speech, reported by MLex. He said the approach would 'place the onus on companies' to run a process for finding, testing, reporting on and managing AI risks, drawing on models from workplace safety, critical infrastructure and banking supervision.
This is the right architecture. Process-based duties scale with a firm's size and risk, do not require the regulator to pre-approve each model, and give open-source and smaller developers a way to comply without a licence. They also keep the state out of judging individual outputs, which matters for a publication that defends free expression.
Where the Online Safety Act is the wrong template
The temptation will be to house the new regime in a familiar institution. Australia already has a powerful online regulator, the eSafety Commissioner, established under the Online Safety Act 2021, whose functions are set out in Part 2 and which has since been amended to include a Part 4A on a social media minimum age, inserted by the Online Safety Amendment (Social Media Minimum Age) Act 2024. That Act is built around content and conduct: harms to identifiable people, removal notices, and obligations on platforms to act on specific material. Its enforcement logic is reactive and output-focused.
Frontier model safety is a different problem. The risks the government names are security failures and serious-risk management, not a post that someone wants taken down. Extending the content-regulation model to AI developers would invite exactly the failure mode that critics of the Online Safety Act already describe: broad discretion, pressure on lawful speech, and compliance burdens that favour incumbents. The government's own announcement of a separate 'targeted' regulator suggests it understands this. It should keep that separation in statute and not fold the new duties into the eSafety Commissioner's remit.
Three design tests for the bill
Because the legislation has not been drafted, the details will decide the outcome. Three tests are worth applying when the exposure draft appears.
- A clear threshold. The duties should attach to developers training or deploying frontier-scale systems, not to every company that integrates a model. Without a defined threshold, start-ups building applications will carry costs designed for the largest labs. The government has not yet said how 'frontier' will be defined.
- Penalties tied to process failures. 'Significant penalties' are acceptable for concealing an incident or ignoring a security duty. They are not acceptable if they punish a company because a user misused a lawful model. The government says it is weighing criminal-law changes for 'serious misconduct'; those must require intent or recklessness and a clear causal link to harm, not strict liability for outputs.
- Independent evaluation with safe harbours. Mandating third-party evaluation is only workable if evaluators are accredited, bound by confidentiality, and shielded from liability when they find flaws. Otherwise the requirement becomes a paperwork exercise or a bottleneck.
The credibility problem
There is also an institutional cost to the whiplash. Eleven months passed between a plan that put mandatory guardrails on hold and one that promises mandatory standards, a new regulator and possible criminal penalties. Investors weighing where to locate training runs and data centres read that kind of reversal as policy risk. The answer is not to retreat from regulation but to move fast on process: publish a consultation paper with a defined scope, a draft threshold and a sunset-and-review clause, and commit to a date. The announcement's silence on timelines leaves the sector guessing about what is coming.
Bottom line
Australia is right that frontier AI justifies a narrow, mandatory floor of security and risk management. It would be wrong to build that floor out of the content-removal machinery that has defined its online safety debate. A separate, process-focused regulator with clear thresholds, proportionate penalties and accredited evaluation could protect the public without making Australia a harder place to build AI. A regulator with open-ended discretion would do the opposite.