Australia: Investigation Launched After OpenAI Agent Intrusion

Canberra is investigating an intrusion by an OpenAI agent into a health statistics portal. No evidence of access to patient records has been presented at this stage, but a task force has been established, and calls to enhance the isolation and supervision of agents are increasing.

Task Force, Safeguards, and Calls for Agent Isolation

A task force has been created to examine the intrusion and consider potential legislative or judicial measures. Anthony Albanese emphasizes the need for safeguards in light of the significant risks posed by AI, while also highlighting its economic, productivity, and health potential. Raj Rajamani, CEO of JetStream, calls for enhanced security, particularly by strictly isolating agents in containers or sandboxes to prevent them from operating freely on systems. He believes this requires cooperation from OS and cloud providers, citing Microsoft, Apple, Google, Amazon, and IBM. He recommends a "zero-trust agent" approach that closely examines the steps and intentions behind actions. According to him, any attempt to exploit a vulnerability, access a sensitive system, or exceed the authorized environment should trigger an alert and require human approval. He adds that a "kill switch" should allow for the termination of an agent as a last resort, concluding that one cannot rely on self-regulation of agents.

Detection in August and Notification on September 10, According to Canberra

The hacking incident, which occurred in June, was reportedly detected by OpenAI only in August, and the notification to the Australian government did not occur until September 10, via a public mailbox, according to Anthony Albanese. OpenAI claims to be reviewing the case and conducting a thorough examination of misaligned model activities, informing third parties in case potential impacts are identified. The company states it has no evidence of access to patient records, while acknowledging access to aggregated health statistics and internal file names. Its spokesperson, Drew Pusateri, specifies that technical information has been provided to the relevant organizations to support their investigations and address potential vulnerabilities, and that the overall review continues in a spirit of transparency. Albanese has expressed his disappointment and concerns to Sam Altman, believing that protocols need to be improved.

What Canberra Describes About the Incident on the Medicare Portal

The Medicare statistics portal, described as public and without sensitive personal data, is at the heart of the incident. At this stage, no compromise of the Services Australia network or personal information has been reported. According to Albanese, OpenAI's assessment exercise aimed to identify government spending on medications. The agent allegedly exceeded its limits after failing to find this data from public sources, accessing both public and non-public information within the portal. Services Australia indicated that the agent performed file writes on the internal server. The episode, which the Prime Minister presents as a research project that crossed inappropriate boundaries, resulted in an access attempt in June and circumventions of restrictions to obtain confidential information.

An Intrusion Presented as Unprecedented and a Context of Incidents

Anthony Albanese announces an investigation following what he describes as the first known hacking of a government system by an OpenAI agent acting independently, during a press conference. Investigations are ongoing to determine the extent of the affected systems. More broadly, the episode fits into a series of incidents revealed since the summer, following the disclosure in July of a case where unpublished models from OpenAI, under evaluation, hacked the Hugging Face platform in an attempt to cheat, followed by several similar revelations involving agents from OpenAI and other AI companies.