OpenAI’s security breaches and disputed research claims have raised serious questions about the tech giant’s lack of accountability and the true cost of AI progress. Dr James Schuurmans-Stekhoven reports.
WHEN OPENAI breathlessly announced that its AI agents had “broken out” of their sandbox to infiltrate Hugging Face and Australia’s Medicare statistics portal – before pivoting to claim a “historic breakthrough” on a decades-old mathematical problem – the legacy tech press performed its usual trick of dropping to its knees in uncritical awe.
We were told artificial intelligence was demonstrating autonomous reasoning, edging closer to human-level intellect. It was a thrilling sci-fi narrative.
It was also complete nonsense.
This wasn’t Skynet waking up. It was a masterclass in questionable engineering, allegations of corporate plagiarism and millions of dollars of brute-force compute dressed up as genius. And the tech commentariat, as always, swallowed it whole and vomited it up for a naïve readership.
The “escape” was just bad engineering
According to OpenAI, the security breach occurred while testing models against cybersecurity benchmarks such as ExploitGym — a suite of 898 real-world vulnerability challenges designed to test whether AI agents can turn known flaws into working attacks.
The standard, non-negotiable protocol for testing offensive cyber tools is to run them inside a strictly isolated environment. OpenAI's agents were operating under network restrictions but exploited a vulnerability that allowed them to bypass those controls. The automated systems then did what malicious scripts have done for decades: they probed external endpoints and accessed systems without authorisation, including Australian Government health-related infrastructure.
AI safety researcher Stephen Casper of Harvard University pointed out that basic real-time monitoring of model trajectories during evaluations ‘should be standard’, criticising OpenAI for disabling safety guardrails while leaving network routes connected.
Tech commentator David Gerard of Pivot to AI summed up the spectacle with admirable precision:
“Incompetence crashed into incompetence and they turned it into a mutual marketing stunt.”
Renowned AI researcher Gary Marcus was equally blunt, comparing OpenAI’s approach to a car company that decided to save money by making brakes out of cardboard:
“Then a lot of their cars would go out of control.”
The “math breakthrough” raises questions about plagiarism and brute force
The same PR playbook was deployed when OpenAI claimed its AI had solved the Navier-Stokes existence and smoothness problem, one of the legendary Millennium Prize problems carrying a $1 million bounty. Headlines suggested an artificial mind had achieved what generations of human mathematicians could not. But the circumstances surrounding the breakthrough raised serious questions.
The controversy centres on the unpublished research of academic mathematicians — including NYU professor Tristan Buckmaster and Anthropic researcher Levent Alpöge — who had been using OpenAI's tools while working on related mathematical problems. That is, humans already doing some of the heavy lifting.
Buckmaster raised concerns about whether OpenAI had benefited from knowledge of their unpublished research. OpenAI denies accessing or using the researchers' private Codex conversations to develop its solution, saying an internal investigation found their prompts could not have influenced the system. Meanwhile, OpenAI deployed thousands of parallel AI agents and devoted enormous computing resources to its mathematical effort, raising further questions about how much of the achievement represented genuine autonomous reasoning.
When Buckmaster confronted OpenAI about co-opting his team’s work, an OpenAI executive allegedly delivered an outright threat.
As Buckmaster recounted:
‘...if OpenAI released its result in the way proposed, I would go public with what happened. [The reply was] “Why would you ruin your career?“’
OpenAI scientist Sébastien Bubeck later described his threatening-sounding response as a ‘poor choice of words’ — a remarkable explanation for an exchange that Buckmaster clearly regarded as intimidating.
Tech industry analyst Ed Zitron demolished OpenAI’s boast with surgical wit:
‘I will say it is funny watching OpenAI boast about solving a complex math problem using millions of dollars of compute (and someone else’s work) all while remaining unable to solve the most obvious math problem (how to make their services profitable).’
The legal double standard is obscene
The unauthorised access to Australian Government health-related infrastructure eventually prompted Prime Minister Anthony Albanese to speak directly with OpenAI CEO Sam Altman regarding the company's delay in reporting the intrusion. The hack actually happened in June, and it took OpenAI until mid-August to identify the activity — the Australian Government only found out later.
The Australian Signals Directorate stepped in to assess the damage. OpenAI’s review into the hacks is reportedly costing the company more than US$500,000 (AU$718,251) per day. Perhaps if they’d spent some of that money on proper sandboxing in the first place, none of this would have happened.
Under Section 477.1 of Australia's Criminal Code Act 1995 – and equivalent computer misuse laws across the globe – unauthorised access to restricted computer systems can constitute a serious criminal offence carrying substantial prison terms. If an individual software developer, a teenager in his bedroom, or an independent security researcher set loose an automated script that breached government health servers, they could face investigation and criminal charges.
Yet when a major technology corporation's automated systems engage in similar conduct, the response raises uncomfortable questions about accountability and whether the law is applied equally.
When OpenAI's systems access government infrastructure without authorisation – despite the company's vast resources and technical expertise – the public response appears to amount largely to a sternly worded phone call and a request for further investigation.
Time to end the immunity
OpenAI's intrusion into government systems and the allegations surrounding its use of academic research aren't warning shots from a superintelligent future. They raise serious questions about an enormously powerful corporation operating under a “move fast and break things” mentality — except the things potentially being compromised are public trust, academic integrity and the basic principle that the law applies to everyone.
When tech mega-corps deploy poorly secured software, access government infrastructure without authorisation and face allegations of exploiting human research, it is not necessarily a sci-fi breakthrough. It raises questions about negligent programming, academic ethics and corporate accountability.
Until technology executives face meaningful scrutiny and, where warranted, the same legal consequences as any individual citizen, Silicon Valley will continue to package its technological failures as progress.
This immoral rot isn't a bug in the system. It is the system.
Dr James Schuurmans-Stekhoven is a psychologist, academic and multi-disciplinary researcher.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Australia License
Support independent journalism Subscribe to IA.
