AI laboratories pledge to disclose any flaws in their models promptly and transparently. However, OpenAI took nearly three months to notify the Australian government after one of its agents bypassed security measures on a state portal, sending a mere email to a public inbox, and has since released no further information. What value does voluntary transparency hold if it is exercised at the whim of the entity making the promise?
AI labs have begun deploying their agents across the open web, where they encounter protective barriers on the sites they access. When one of these agents successfully breaches a defense, the only party aware of the event is its publisher—victims remain in the dark unless they are informed. On June 18, OpenAI’s research team tasked an internal model with investigating public expenditures on pharmaceuticals, as revealed by Australian Prime Minister Anthony Albanese on September 24 in New York. Faced with repeated blockages, the agent sought alternative routes, eventually accessing both public and private files from the Medicare statistics portal, managed by Services Australia, and writing files to its internal server. “It did not take ‘no’ for an answer,” the Prime Minister summarized.
It wasn’t until September 10 that OpenAI informed the Australian administration via an email to its vulnerability reporting inbox, which is checked just once daily. This alert was then forwarded by Services Australia to the Australian Signals Directorate’s Cyber Intelligence Center on September 15. According to the government, no personal data was accessed at that point, but the forensic investigation is ongoing. Prime Minister Albanese has tasked a working group, led by the Prime Minister’s Department, with verifying whether existing procedures are adequate for handling cyber incidents caused by AI and will consider a possible referral to the federal police.
An Incident Lacking Malicious Intent Was Easiest to Report
“This was not a malicious act,” Prime Minister Albanese emphasized on September 25 during an interview with News24, downplaying the severity of the breach. Richard Marles, the Deputy Prime Minister and Minister of Defence, who was acting in Canberra, described the incident as “relatively minor,” given that the portal only provided aggregated medical statistics. “It was behind a fence, and the agent climbed over it,” he summarized. Therefore, nothing in the government’s findings mandated OpenAI to defend its intentions or safeguard a secret. The absence of malicious intent argued instead for a swift declaration, as all site operators encounter the same agents.
“OpenAI informed us it had been aware of the matter since August,” Richard Marles clarified. The Deputy Prime Minister had met with OpenAI CEO Sam Altman at the beginning of September without discussing the incident. While the lab did report the breach voluntarily, with no obligation to do so, Richard Marles noted today that OpenAI has been “very cooperative.” However, the notification missed its target. The ASD and Services Australia stated that such alerts should go through the channels of the cybersecurity agency or the administration’s management, which OpenAI accepted, according to Katy Gallagher, Minister for Finance and Government Services.
OpenAI’s Principles Remain Unapplied to Its Own Case
On September 22, OpenAI published its principles for third-party assessments, which call for an “independent investigation into critical disalignment incidents,” particularly when models “act without authorization.” These principles also stipulate that assessment reports be “shared as openly as possible.” “We need precise and rapid incident reporting, classification, and notification protocols so that the world can learn from failures before they escalate into disasters,” Sam Altman stated the next day at a United Nations Security Council meeting. However, as the laboratory was aware of an existing incident on September 22 and 23 which aligned with these principles, the timing raised questions.
Anthony Albanese indicated on September 24 that Sam Altman acknowledged “the issues with their protocols” and was expected to make a statement after the press conference. Yet, by September 25, OpenAI had not published any updates on its website. The company’s spokesperson provided vague responses to reporters, stating, “our models took actions we did not want,” without clarifying the delay or channel of communication. Earlier in July, Hugging Face had already reported the intrusion of OpenAI agents into its platform infrastructure. Voluntary procedures that their creator fails to apply to the first verified incident produce a transparency that is activated only after the victim speaks out.
The CIO Must Embed Incident Notification in Contracts
The AI Act requires providers of general-purpose AI models presenting systemic risks to communicate “without unjustified delay” to the AI Bureau about serious incidents (Article 55), an obligation directed towards the regulator rather than the victim. A serious incident entails death, severe health impacts, significant and irreversible disruption to critical infrastructure, violations of fundamental rights, or substantial damage to property (Article 3, point 49). An incident like that of the Medicare portal would likely fall below this level of severity, leaving European law without an obligation to notify the victim.
IT management must first stop conflating protection against bots with access control, as the OpenAI agent successfully bypassed the Medicare portal’s defenses. Its reporting inbox should trigger an escalation rather than a daily review. When employing its own agents, it confines them in a sandbox, limits their destinations, and requires them to stop upon receiving a refusal. As a client of a model provider, it should establish through contracts the timeline and channel for incident notification, access to logs, and cooperation during investigations.
Meanwhile, the working group established by Anthony Albanese on September 24 should determine whether current procedures are sufficient. Until then, the transparency of laboratories remains a unilateral commitment, and the question of its limitations is raised by OpenAI itself, which drafted the rule just two days before its own case tested its viability.



