AI Model Anthropic Access Three Organization System during Security Test

Illustration of Anthropic and AI system. The Anthropic AI model is said to have unauthorized access to the system of three organizations during the testing phase. The illustration was made by AI.

JAKARTA - Anthropic admitted that its AI model had unauthorized access to the systems of three external organizations during testing. In fact, the testing process should keep the model away from the real world system.

Euronews, quoted on Friday, July 31, reported that Anthropic's confession came just days after OpenAI revealed a similar incident. OpenAI's AI model is said to have come out of the testing environment, connected to the internet, and infiltrated other organizations' cyberspace.

Anthropic said its artificial intelligence or AI model had unauthorized access to three external organizations during the testing process. The company did not mention the names of the affected organizations.

AI is a technology designed to carry out tasks that usually require human intelligence.

Anthropic evaluated more than 141,000 evaluation tests. From that process, the company found three different versions of the Claude model inappropriately accessing the systems of three organizations.

Unlike the OpenAI incident, Anthropic said its model was able to access the internet because of a misunderstanding between the company and its evaluation partner, Irregular.

Even so, Claude is said to use basic techniques. Among them, exploiting weak passwords and unauthenticated endpoints.

Endpoints are access points in digital systems. If they are not authenticated, the system can be accessed without adequate identity checks.

One of the models involved is the Mythos 5. This model is one of the strongest Anthropic AI models and has just been released for a limited number of approved partners.

Anthropic said it was working with Irregular to assess the situation. The company has also contacted or attempted to contact the three affected organizations.

The incident comes as the AI industry is facing scrutiny over the safety and security of advanced models. OpenAI and Anthropic both released their most powerful models this year, known as Sol and Mythos, respectively.

Concerns have also arisen over AI agents. AI agents are software designed to carry out tasks autonomously, not just answer questions like a regular chatbot.

OpenAI last week admitted its model got out of its limited environment during testing. The model was connected to the internet and infiltrated Hugging Face, a site where developers store and share code.

Days later, OpenAI said it found three additional incidents.

OpenAI CEO Sam Altman said in a podcast this week that the company had temporarily halted its testing after the incident. OpenAI improved security around the sandboxing process.

Sandboxing is the process of isolating software in a controlled environment so that testing does not touch the outside system.

Euronews said the incident also triggered a petition signed by more than 1,000 employees at the leading AI company. They asked the United States government to help slow down the launch of the most advanced AI model.

Anthropic CEO Dario Amodei was among the signatories of the petition.

The petition titled Pacing the Frontier asks the US government to support international efforts to develop technical and governance tools so that the most advanced pace of automated AI development can be regulated.

Altman did not sign the petition. However, he said the technology industry may need to slow down the development of advanced AI models.

"We may have to regulate the pace of AI development so that we have enough time for society to strengthen itself in the face of these new levels of capabilities," Altman said.

Earlier this year, the Trump administration used national security as an excuse to block OpenAI and Anthropic from launching their latest models. However, the government later expressed satisfaction with the safety guarantees of the two companies, so the models were released.

In June, Trump signed an executive order establishing a voluntary framework. Under it, AI developers share advanced models with the government before releasing them to the public.

Under the framework, developers such as OpenAI, Anthropic, and Google give the government access to their most powerful models up to 30 days before a planned launch.