Australia AI Kill Switch and Data Centre Control Bill Guide

Australia is moving from voluntary AI and data centre expectations towards legislation, and two very different proposals are now on the table: a crossbench bill with a hard AI kill switch and a data centre moratorium, and a government plan to legislate national AI standards in 2027. On 7 September 2026 independent MP Andrew Gee introduced the AI Kill Switch and Data Centre Control Bill 2026 in the House of Representatives. Two months earlier, on 15 July 2026, the Prime Minister announced that the government would legislate Australian Standards for AI and a mandatory framework for large data centres. Neither is law. This guide explains what the Gee bill would require, how it differs from the government's plan and the US AI Kill Switch Act, what its chances are, and what AI providers, data centre operators and security teams can usefully do now. It is written for compliance, AI platform, infrastructure and security leaders with Australian operations.

Source note: the official bill page, bill text and explanatory memorandum on aph.gov.au could not be retrieved when this guide was written. Bill details come from Policai's instrument note, which cites the bill as introduced section by section. Digital Policy Alert and Michael West Media confirm the bill, its sponsor, its introduction date and its broad subjects, but not its penalty amounts, deadlines or thresholds, so those figures are omitted here. Status is as of November 2026. Confirm against the official text before relying on it.

Key takeaways

It is a private member's bill, not law. The AI Kill Switch and Data Centre Control Bill 2026 (APH bill ID r7537) was introduced and its second reading moved on 7 September 2026; it remains before the House, and commentators expect it not to pass.
Part 2 is a kill switch for large AI systems. Providers of AI systems above a compute-cost threshold set by rules would have to be able, at all times, to stop inference, suspend user access and shut the system down.
Incident clocks are tight. Critical incidents, including sabotage of a lawful shutdown and loss-of-control scenarios, would need prompt notice to the Minister and a follow-up written report, according to Policai's summary.
Part 3 is a data centre moratorium. Constitutional corporations could not build or upgrade AI data centres (defined by power capacity and high-density or liquid cooling design) until Parliament legislates a set of safeguards, according to Policai's summary.
The government's track is different. It plans national AI standards and a mandatory data centre framework, with legislation expected in early 2027; capability built for a kill switch and incident reporting helps under either.

Status and timeline

Date	Event	Source
15 July 2026	Prime Minister Anthony Albanese announces Australian Standards for AI, a mandatory framework for large data centres and copyright protections; Office of AI established in the Department of the Prime Minister and Cabinet	ALP media statement; ACS Information Age
August 2026	National Cabinet to consider the government's approach	ALP media statement
7 September 2026	Andrew Gee MP introduces the AI Kill Switch and Data Centre Control Bill 2026; second reading moved the same day	Policai, citing the APH bill page; Digital Policy Alert
26 September 2026	Michael West Media reports the bill has support from independent Dai Le and calls it unlikely to pass	Michael West Media
23 October 2026	Bill recorded as "Before Reps"	Policai
Early 2027	Government legislation on AI standards expected	ALP media statement; ACS Information Age

Private members' bills in the House rarely proceed to a vote without government support. If the bill did pass, Policai reports that the whole Act would commence the day after Royal Assent (s 2). Treat everything below as what the bill would do, not what the law requires.

Part 2: the kill switch

Part 2 applies to a "covered technology": an AI system developed using computing power whose cost exceeds an amount prescribed by rules (s 7). The responsible providers are constitutional corporations that make such systems available to others through a programmatic interface, hosted service or similar mechanism, other than solely for non-commercial purposes (s 8). Until the Minister makes rules setting the compute-cost amount, no system is covered, so Part 2's duties would have nothing to attach to on day one.

Duty	What the bill would require	Civil penalty (as reported)
Shutdown capability (s 12)	At all times be able to stop inference, suspend user access and shut the technology down	Civil penalty, each day a separate contravention (amount not confirmed by a second source)
Critical incidents (ss 9 to 11)	Defined to include sabotage of a lawful shutdown attempt, unintended operation causing death or injury, and a loss-of-control scenario	Not applicable (definitions)
Incident notification (s 16)	Notify the Minister quickly, then provide a written report (deadlines not confirmed by a second source)	Not reported in sources checked
Emergency directions (ss 17 and 18)	Comply with a ministerial direction, including to shut down, where a critical incident presents a significant risk of serious harm or damage	Civil penalty for failure to comply (amount not confirmed by a second source)

The bill does not name an administering department; it confers powers on "the Minister". Michael West Media notes it does not address criminal responsibility of operators or developers. Digital Policy Alert also classifies the bill as containing human oversight, testing, performance monitoring and design requirements, but its public summaries do not describe their content, so they are not detailed here.

Part 3: the AI data centre moratorium

From commencement, a constitutional corporation must not construct a new AI data centre or upgrade an existing one, backed by a civil penalty (ss 22 and 23). An AI data centre is defined by power capacity and by design for high rack density or liquid cooling (s 21); Policai gives the thresholds, which we could not confirm in a second source. "Upgrading" is not defined, so its reach over existing sites is uncertain.

The moratorium ends only when other Acts ensure a list of conditions and an Act expressly ends Part 3 (s 24). The conditions reported include that data centres do not raise consumers' water or power bills, that affected communities can approve or reject a data centre, and that the Commonwealth reviews and approves AI systems before they are first released or supplied in Australia. The last condition would tie data centre construction to a national pre-market approval regime for AI, which Australia does not have.

How the proposals compare

Feature	Gee bill (Australia)	Government plan (Australia)	AI Kill Switch Act (United States)
Status	Private member's bill before the House	Announced policy; legislation expected early 2027	Bill introduced in the House on 23 July 2026
Who is covered	Providers above a compute-cost threshold set by rules; any corporation building AI data centres	Large data centres; AI standards scope not yet published	Frontier providers above compute-cost and revenue thresholds
Shutdown duty	Stop inference, suspend access, shut down at all times	None announced	Mandatory shutdown capabilities and emergency orders
Incident reporting	Prompt notice to the Minister, then a written report	None announced	15 day incident clock in the bill
Data centres	Moratorium until listed conditions are legislated	Fund new power supply, pay grid connection costs, reduce load when needed, maximise water efficiency	Not covered
Regulator	"The Minister"; no department named	Office of AI in PM&C to drive implementation	Department of Homeland Security

For the US bill in detail, see the AI Kill Switch Act Guide. For Australia's existing voluntary AI guardrails, see the Australia AI Guardrails Guide.

Why it matters for security and identity teams

A legal kill switch is only as good as the operational ability behind it. "Stop inference, suspend user access and shut down" translates into concrete controls: the ability to revoke every API key, OAuth client and service token that reaches the model; to disable agent identities that call it; to drain and stop serving infrastructure; and to prove afterwards that it worked. That is identity and access engineering, not only model engineering.

Access suspension: requires a complete inventory of API keys, tokens and client credentials issued to customers and internal agents, and a tested way to revoke them in bulk. See the AI Agent Kill Switch and Containment Guide.
Loss-of-control incidents: agent behaviour that escapes its intended scope is a live risk, not theory. The OpenAI Agent Medicare Portal Breach 2026 shows an AI agent reaching non-public Australian government data, and Michael West Media cites it in coverage of the bill.
Sabotage of shutdown: the bill treats interference with a lawful shutdown as a critical incident. Shutdown controls must sit outside the reach of the systems they stop, with separate administrative identities and break-glass access.
Data centre physical and logical security: any mandatory framework will sit alongside SOCI Act obligations for data storage and processing assets. See the Australia SOCI Act Guide and the Data Centre Security Guide.

What to build now: a readiness programme

NHI Mgmt Group recommendation: this section and the decision guidance below are our own guidance, not requirements of the source document.

None of this is required yet. These steps are worth doing because they reduce real risk and because both Australian proposals and the US bill point the same way.

Map exposure. List AI systems you provide to others through APIs or hosted services, their training compute (or the provider's), and any Australian data centre build or expansion plans. Flag large sites and any designed for high-density racks or liquid cooling.
Define the shutdown levels. Document three actions separately: stop inference, suspend user access, full shutdown. For each, name the owner, the systems touched, the identities that must be revoked and the expected time to complete.
Build bulk revocation. Make sure every key, token and agent identity that reaches the model can be disabled from one control point, and that the control point is not reachable by the model or agents.
Test in a non-production environment first. Run the shutdown against a staging deployment, measure time to full effect, and confirm restore works. Then run a timed exercise against a low-risk production model with customer notice. Roll back by re-enabling credentials from a pre-recorded state.
Set up a fast incident path. Define what counts as a critical incident (harm, loss of control, shutdown interference), who decides, and a draft notification template, so a regulator clock could be met if one is legislated.
Track the legislation. Assign someone to watch the bill's progress, the Office of AI's standards work and the data centre framework, and re-assess when exposure drafts appear.

Decision guidance

Frontier or large model providers serving Australia: build shutdown and incident capability now; it is likely to be expected under any final regime and is already expected in the US bill.
Organisations consuming AI through APIs: you are not a responsible provider under Part 2 as described, but you need your own ability to cut off agents and integrations if a provider shuts down or is directed to.
Data centre developers: the Gee bill's moratorium is the highest-impact provision but the least likely to pass in its current form; the government's framework (power, grid cost, water) is the more probable obligation. Plan for that.
Small organisations: no direct duty; focus on supplier contracts that require notice if an AI service is suspended or shut down.

Detection and measurement

Time to suspend all access: measured in exercises, from decision to the last credential revoked.
Credential coverage: percentage of keys, tokens and agent identities reachable by the bulk revocation control.
Shutdown control isolation: number of identities with rights to alter shutdown controls, kept small and reviewed.
Incident triage time: time from detection of a potential loss-of-control event to a documented critical or not critical decision.
Legislative watch freshness: date of the last status check against the APH bill page and Office of AI publications.

Log every use of shutdown, suspension and bulk revocation controls, every change to who can use them, and agent actions outside approved scope. Mapped to NIST SP 800-53 Rev 5, this sits under IR-4 (incident handling), AC-2 (account management) and AU-9 (protection of audit information); in NIST CSF 2.0, RS.MI-01 (incidents are contained) and GV.RR-01 (leadership accountability).

Common failure patterns

Treating a bill as law, or ignoring it entirely. Both are mistakes. Plan for direction of travel, not exact text.
A kill switch that only stops the model. If agents and integrations keep valid tokens, access is not suspended.
Shutdown controls reachable by the system they stop. The bill's sabotage incident category exists for this reason.
No owner for regulatory incident clocks. A short notice deadline is impossible without a pre-agreed decision path.
Data centre plans that ignore policy risk. Projects designed without the government's power and water expectations in mind may face delays under either track.

Practitioner checklist

Record the bill's status and re-check it monthly against the APH bill page.
Identify AI services you provide through APIs or hosted services in Australia.
Document stop inference, suspend access and full shutdown procedures with owners.
Build and test bulk revocation of keys, tokens and agent identities.
Isolate shutdown controls from the systems they control.
Draft a fast critical incident notification process.
Flag large data centre projects and any designed for high-density or liquid cooling.
Track Office of AI standards and the data centre framework through 2027.

Standards and references

Related NHI Mgmt Group resources: AI Kill Switch Act Guide · Australia AI Guardrails Guide · AI Agent Kill Switch and Containment Guide · Australia Cyber Security Act 2024 Guide. This guide summarises a bill and government announcements and is not legal advice.

Australia AI and Data Centre Laws Guide: The AI Kill Switch and Data Centre Control Bill