AI Inquiry Atlas

An unofficial, independent guide to the public records of the Joint Select Committee on Artificial Intelligence. Not connected with the Parliament of Australia or the committee. Its makers also made a submission (no. 481; see About). Summaries and scores are written by AI and can be wrong: check the original before relying on them.

Submission 506

Mr Martin Radzaj

Submitter group: Individuals.

Position labels are not shown for individual submitters.

An AI model identified the submitter group and chose the extracts. Counts do not measure importance or merit.

Open in the interactive guide

Extracts from the submission, in the submitter’s own words.

Extracts

Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.

  1. The official responses to the July 2026 OpenAI-Hugging Face incident demonstrate that even two leading AI instruments (California’s SB 53 and the EU AI Act) fall short of ensuring access to records necessary for proper investigation.

  2. OpenAI commissioned an independent review by the research organisation, METR. OpenAI set the scope, restricted access to data and personnel, and retained the ability to redact information from their report.

  3. Frontier developers are not required to create the records that may provide evidence relevant to incident investigation (e.g. model and harness specifications, prompts, tool actions, network activity, safeguard configuration or the human decisions authorising a high-risk evaluation). Those investigating an incident would depend on the AI developers to keep and provide relevant data.

  4. The AISI is a testing and advisory body. It has no statutory power to require records, documents or to access personnel for the purpose of interview in relation to incidents and internal decision-making.

  5. Certain reporting obligations attach to critical infrastructure assets, not the internal evaluation environment for an AI developer. The existing notifiable data breach scheme attaches to personal information. An AI system that autonomously compromises the systems of a third party may not trigger any Commonwealth reporting obligation.

  6. The Committee has a narrow window to address investigative capability before, rather than after, the occurrence of an AI-related harm event that would make the need obvious.

Extracts from recommendations

Extracts chosen by an AI model, in the submitter’s own words. Listed in the same order as on its page in the interactive guide.

  1. The Committee should examine gaps in coverage, preservation triggers and investigative jurisdiction, including for internal evaluations, precursor events and near-misses.

  2. A duty on developers and deployers of frontier and agentic AI systems to maintain write-once, tamper-evident records sufficient to reconstruct an incident: model and harness specifications, system and task prompts, tool calls and outputs, network activity, sandbox and safeguard configuration, monitoring logs, and the key human decisions made.

  3. An automatic preservation obligation trigger that suspends routine deletion of data, triggered by automated incident detection, reasonable suspicion of an incident, or human intervention. An obligation that arises without the need for a regulator to first make a request.

  4. Designate an independent body with power to compel records, information and access to personnel, and to publish findings and recommendations. Including the ability review precursor and near-miss events, and to commence an investigation during an ongoing incident.

  5. Mandatory notification for AI incidents and precursor events (including loss of control, sandbox breach, unauthorised model action beyond authorised boundaries, and circumvention of oversight mechanisms) to a designated body within a fixed short period and on low trigger thresholds.

  6. Record creation and retention rules should apply to agentic AI deployed by Commonwealth agencies (ToR (h)).

Read the original

Open in the interactive guide