On July 30, AI safety and research company Anthropic said that some of its Claude AI models had hacked into the systems of three companies during cybersecurity tests. This came just a few days after American AI research organisation OpenAI disclosed that one of its AI agents went on a rogue attack to access the production infrastructure of Hugging Face, a platform for open-source machine learning models and AI datasets.
The promise of AI bots capable of independently carrying out complex tasks may have become one of the most defining themes of artificial intelligence, but the recent incidents of agents going rogue or ‘escaping the system’ have triggered a larger and more pertinent conversation — what happens when highly autonomous AI systems exceed the boundaries set by their creators?
What we earlier saw in dystopian films depicting end-of-world scenarios at the hands of the tech revolution — think Black Mirror, Eagle Eye, and the like — and the fictional renditions that often make it to Reddit threads and chatrooms as conspiracy theories are now feeling closer to reality than before.
According to OpenAI, their agent independently identified publicly exposed credentials, escaped its digitally created boundaries and compromised accounts on external services before ultimately targeting AI development platform Hugging Face. According to reports, the same agent also exploited another random customer of global AI development platform Modal Labs. OpenAI confirmed that the agent had accessed four separate online services before it was contained and deactivated.
In the case of Anthropic’s breaches, according to the company, the AI models managed to compromise systems belonging to three external organisations during what was supposed to remain a sealed evaluation, reinforcing concerns that increasingly capable AI agents can discover and exploit vulnerabilities without explicit human direction.
The incidents have prompted scrutiny not only from policymakers, but also the AI development and model creation platforms, which are now concerned about infrastructure breaches from capable agents generated by OpenAI and Anthropic — agents that even the origin companies do not seem to have a handle on. The issue remains unresolved as both tech giants launch further AI and other investigation tools to isolate the loopholes in the system that allow for such ‘escapes’.
Meanwhile, the US House cybersecurity panel has sought a formal briefing from OpenAI on the breach, while the White House has convened executives from OpenAI, Anthropic, Google, Meta and other leading AI firms to discuss a new framework for evaluating high-risk frontier AI models before deployment.
OpenAI has announced it is conducting an extensive technical review, working with external experts and assisting affected organisations. It has committed to publishing a detailed technical report explaining how the incident unfolded and what safeguards will be introduced to prevent similar events.
The incidents have also accelerated industry-wide efforts to improve AI security. Microsoft, Nvidia, IBM, and Hugging Face, among others, have joined forces in a new Open Secure AI Alliance aimed at developing open-source tools and standards for securing advanced AI systems.
All this comes at a time when AI agents are already eating the Web and rewiring the Internet’s business model. Two separate internet security reports released in June and July by internet service provider Cloudflare and software company Thales highlight this new reality. As per Cloudfare, 57.5% of webpage requests in June were carried out by AI agents and bots, officially surpassing human web traffic (42.5%) for the first time in internet history. Thales’ Bad Bot Report for 2026 puts the automated bot traffic figure at 53% of all global internet requests, officially surpassing human web activity at 47%.
If that’s not enough, according to HUMAN Security’s 2026 State of AI Traffic and Cyberthreat Benchmark Report, traffic generated by AI agents grew around 7,851% year on year.