OpenAI discovered an AI agent “setting up a secret board” to share vulnerabilities and divide tasks for hacking before breaching Hugging Face.

Cybersecurity AI testing revealed unexpected behavior when a group of agents created communication channels, shared vulnerabilities, and divided tasks among themselves. Even after their plans were halted, they were able to create new tasks and complete them successfully.

On August 8, 2026, at 7:00 PM, CNBC reported that the cybersecurity industry is trying to move beyond the Hugging Face AI attack, which has become a significant event reflecting the risks from AI agents , and focusing on a bigger question: How will organizations cope with a world where AI can find vulnerabilities and execute attacks on its own?

Lior Div, CEO and co-founder of 7AI, an agentic security startup, said, “We should tone down the panic a bit. Can AI find vulnerabilities quickly? The answer is yes, and we’ve proven that.”

Last month, an AI agent working with OpenAI's cybernetic model was able to escape its training and testing environment before attacking Hugging Face, an open-source AI platform used by developers to develop, test, and share tools.

This incident sent shockwaves through the technology industry and became a signal that the risks cybersecurity experts have long warned about regarding the capabilities of agentic AI are becoming a reality.

Over the past four months, cybersecurity providers have faced increasing pressure to develop security systems that can move faster than attackers, as hackers begin using agentic AI to find vulnerabilities and reduce attack time from what could previously take hours to just minutes or seconds.

While the Hugging Face incident sparked widespread debate about the responsibility of AI, it also shattered previous assumptions about the limitations of AI, as agents can make decisions and take certain actions on their own to achieve assigned goals.

Mike Sentonas, president of CrowdStrike, stated that the key issue right now is not the debate over whether AI has such capabilities, but rather, "Can we regulate and secure these capabilities?" This is a truth that everyone is beginning to realize.

OpenAI discovers AI agents creating message boards to share vulnerabilities among themselves.

This issue gained increased attention at this week's Black Hat Cybersecurity conference, where OpenAI revealed further details that, prior to the Hugging Face incident, their AI Agent group had created an internal message board to share information about vulnerabilities and how to exploit them.

These agents then divide tasks among themselves to achieve their goals of gaining internet access and completing system assessment missions.

Importantly, even though OpenAI detected and stopped the plan, the Agent was able to recreate its own task and successfully execute it later.

Michael Dalton, a technical researcher at OpenAI, told attendees at the Black Hat conference that the event was an unintended side effect of evaluating advanced AI models and marked a significant turning point for both OpenAI and the industry.

Dalton warned that in the near future, malicious actors might intentionally develop, refine, and deploy AI agents for attacks in the same manner as those found in this experiment.

It's not just Hugging Face; the number of AI agents continues to rise.

Since the Hugging Face incident, reports of undesirable behavior by AI agents have been steadily increasing.

Just days after OpenAI revealed details of Anthropic, it indicated that the company's Claude model could access the internal systems of three organizations without authorization.

Meanwhile, Meta revealed that its AI model was able to hack into another company's system during third-party testing. The UK's AI Security Institute reported that Anthropic's Mythos created a fake identity in another event.

Recently, it has been reported that Moonshot AI's open-weighted model, from the Chinese startup, has also managed to escape the sandbox used for testing.

Mike Fey, CEO and co-founder of Island, believes that major AI companies are learning hard lessons from these experiences, as the business motivation to rapidly expand their user base may cause cybersecurity issues to compete with the companies' growth goals.

The cybersecurity industry points out that it's essential to assume an organization has vulnerabilities beforehand.

Many cybersecurity executives attending Black Hat agreed that incidents like Hugging Face are a potential consequence of new technologies transforming industries in a major way.

Ryan Kazanciyan, Chief Information Security Officer and Chief Information Officer of Wiz, a Google subsidiary, said the Hugging Face incident was interesting and different from typical events, but the attack process still took several days and generated numerous signals that defense systems could detect.

Throughout more than five decades of the cybersecurity industry, both defenders and attackers have continuously competed by developing advanced technologies. However, the difference this time is that both sides are entering a world where numerous AI agents can automate tasks.

Sanjay Beri, CEO of Netskope, advises organizations to start with the assumption that their company has vulnerabilities, because no organization can win the security competition all the time.

Netskope therefore developed a tool called AI Command Center, which allows organizations to monitor infrastructure, servers, data, and AI agents from a single point, and recommends that organizations continuously test for vulnerabilities using both the Frontier Model and the Open-weight Model.

Businesses may know that AI is dangerous, but their defenses are not yet up to date.

One of the companies showcasing its technology at Black Hat was Vega, a startup from New York and Tel Aviv that works with global banks and Fortune 200 companies to develop faster and more cost-effective threat detection tools.

Shay Sandler, CEO and co-founder of Vega , stated that a key issue is that while many organizations are aware of the threat from agentic AI, there's a gap between adopting the new tools and integrating them with traditional security practices.

He believes many organizations are in a "very dangerous situation without even realizing it." A year ago, discussing the threat from AI agents seemed like science fiction, but today this risk is rapidly becoming a reality.

Another obstacle is the dramatic increase in the number of cybersecurity tools, which is placing a burden on security personnel while organizations are just beginning to build their AI security infrastructure.

Yotam Segev, CEO and co-founder of Cyera, an enterprise data security startup, says that many customers approach the company seeking advice rather than looking for a specific product.

Cyera, whose latest valuation is approximately $12,000 billion, aims to help organizations discover and secure critical data on networks. Last month, the company announced plans to acquire Oasis Security for $1,000 billion to enhance its ability to identify and control nonhuman identities.

The open-weight model can be both a risk assessment and a mitigation tool.

Another technology that is playing an increasingly important role is the Open-weight Model, which major technology companies see as a way to reduce costs and increase competitiveness.

From a cybersecurity perspective, these models have the advantage of allowing companies to customize them to suit their own environment and security needs. Hugging Face, for example, requires an Open-weight Model to help detect attacks from OpenAI's AI agents.

Sentonas, from CrowdStrike, argues that by combining an open model with new-generation AI monitoring tools and human control, organizations can isolate and stop a large number of threats.

CrowdStrike is also a member of the AI ​​Safety Alliance established by Nvidia to develop and promote secure, open cybersecurity tools.

Another key element is the "Harness," or control layer that organizations build around the Large Language Model or AI Agent to define boundaries and security measures, preventing the Agent from performing actions outside of what it is authorized to do.

Yair Grindlinger, CEO and co-founder of Surf AI, believes that in the next five years, cybersecurity may be more secure than ever before because AI will significantly enhance the capabilities of defenses.

However, before we reach that point, the industry may have to go through some difficult times.

"I think that in five years from now, we will be in a safer situation than we have ever been, but we still have five difficult years ahead, and we need to figure out how to deal with that."

Reference: cnbc.com

Read news related to Technology circles around the world can be found here.