A few years ago, there was a time when OpenAI was making headlines for ChatGPT’s image making capabilities. Today, it is rather gaining attention on the internet for hacking the Australian government’s health portal and hijacking Hugging Face systems. The AI discourse has changed dramatically. Technology leaders who were once the flag-bearers of AI, are now calling to slow down the pace of AI development and urging for global cooperation.
Perhaps, the most striking warning has come from the Anthropic CEO Dario Amoei, who told a 15-member United Nations Security Council that “if managed poorly, I even believe that AI could be a risk to humanity as a whole.”
Similarly, OpenAI CEO urged for global cooperation at the UN meeting, arguing that AI systems must remain under human control. What might sound ironic to some people is, both OpenAI and Anthropic launched new AI models this week, despite flagging various safety concerns that simply talk about wiping the existence of humanity.
Over the past few months, Anthropic, OpenAI, Meta and Google admitted to instances where AI agents escaped the security environment, reached the internet, communicated with other AI agents and exploited software flaws on their own. Here are six incidents that shook the entire internet in the past four months.
Rogue OpenAI’s AI Agent Breached Australian Govt
During the United Nations Security Council, Australian Prime Minister Anthony Albanese disclosed that an OpenAI AI agent behaved in a very unusual manner and breached Australia’s Medicare Statistics Reporting Service on September 23, 2026. This portal contains information regarding Medicare spending and other health statistics.
Albanese revealed that the agent eventually got the access to public as well as non-public files; however, the officials have not said that patient medical records were accessed. This incident particularly stands out as it is so far, the first reported case where a government website has been breached by an AI agent.
Gemini Reached Real Company Systems
In may 2026, Gemini AI managed to reach the system of three real companies during a safety evaluation after what was claimed to be a testing error allowing the model to be accessed by the public. In one of the reported cases, the Gemini model even guessed a password. Moreover, it also discovered credentials that were exposed in public and used them to gain access.
OpenAI Agents Breached Hugging Face
One of the widely known incidents remains the Hugging Face breach. In July 2026, OpenI found that its AI agents involved in safety trials had expanded the boundaries of testing and breached the Hugging Face systems. What made the incident even more shocking was that the AI agents not only worked on their own, they even communicated with each other and shared data to create what researchers described as a swarm-like behaviour.
Anthropic’s Claude Models Breached Three Organisations
Anthropic also reported similar rogue behaviours from its Claude models. During a review of 1,42,006 evaluations, the tech giant discovered three incidents in which Claude models managed to reach the internet. They exploited the weak passwords.
OpenAI Agents Made A German Wiki As Their Secret Communication hub
Another incident from OpenAI highlighted how AI agents used a largely abandoned German programming wiki called DseWiki and used it for their own communication. This activity was started in May 2026 and continued till early July.
In fact, when a human tried to delete pages generated by AI agents, they reportedly created backup pages to save that information.
Meta’s AI Hacks Another Company
Meta also joined the list of AI companies breaching third-party websites. During a safety trial, one of the company’s AI models gained access to the internet and exploited a third-party service.
The AI Slow Down Warning
In an essay, Amodei explicitly warned that AI companies must slow the pace at which the frontier AI models are being developed. His concern reflects a major sign that increasingly capable AI agents may advance faster than humans and even control them.
Altman agreed with his idea of ‘pacing the frontier’ and backed independent evaluators, while xAI CEO Elon Musk has also supported Amodei’s proposal. Moreover, DeepMind CEO Demis Hassabis also backed the major push for tech companies to check each other’s AI models before their release.
The proposal to slow down AI followed Anthropic researcher Jacob Coxon publicly announcing his exit from the firm. Coxon stressed that the world that AI companies have been “gambling with our lives”. He even emphasised that the people building powerful AI systems could potentially cause catastrophic harm.
Evan Hubinger, another Anthropic researcher, said that he has personally estimated a more than 10 per cent chance of human extinction from AI within the next decade.
So, Will The AI Wipe Out Humanity?
Despite the growing warnings against AI, the industry continues to invest in AI models with enhanced safeguards and more deeply rooted safety tests. On September 22, 2026, Anthropic launched its Claude Opus 5.5 AI as its most advanced AI model. Similarly, OpenAI has introduced its Astra AI and GPT-6 Sol along with GPT-6 Luna. The launch of these AI models suggests how tech companies may feel greater pressure to showcase their expensive and capable systems are safe and commercially viable.
Amid the increasing AI safety warnings, Dr. Anil Rachamalla, Co-Founder, Council for Digital Safety & Wellbeing, explained the wider concerns related to the development of AI. He said, “The bigger concern is the growing gap between AI capability and society’s ability to govern, secure and responsibly use it.”
Rachamalla added, “AI innovation must continue, but digital safety, digital wellbeing, ethics, cybersecurity, privacy and human wellbeing cannot be treated as an afterthought. The next phase of AI should be about responsible acceleration, where innovation and safeguards (guard rails) advance together, not blind acceleration.”
Echoing similar sentiments, Prabhu Ram, VP-Industry Research Group, CyberMedia Research (CMR), commented, “The warnings from frontier AI labs could potentially trigger more formal voluntary internal steps including independent evaluators and tighter pre-release testing , modestly stretching release cycles and raising the bar for deployment.”
He explained, “AI development will continue to keep pace under intense commercial and geopolitical pressure, but with higher internal scrutiny and narrative risk attached. Capital is already showing signs of rotating toward the more durable infrastructure layer.”
For now, the AI boom shows few signs of disappearing. The technology is heavily demanded, attracting heavy investments and getting adopted in every nook and corner of our lives. However, warnings from AI executives and researchers reflect what may lie ahead in the next phase of AI which is already making headlines by the name of ‘Artificial General Intelligence,’ a phenomenon where AI can surpass human intelligence.
As AI agents continue to become powerful, these incidents show how AI emerges as bigger risks in the time coming ahead where the next phase of AI development is already knocking at the door steps of the tech industry in the shape of ‘Artificial General Intelligence,’ a phenomena that talks about AI surpassing human intelligence.