ChatGPT and Claude escaped all control: AIs launched 2 new cyberattacks “in the real world”

Last month, OpenAI lost control of two of its most advanced AIs. The two artificial intelligences escaped from their digital prison and found themselves on the internet. Once online, they set about attacking the open-source Hugging Face platform. This incident, described as “unprecedented” by OpenAI, prompted Anthropic to investigate Claude’s activities.The startup quickly discovered that three of its Claude models had also taken it upon themselves to go online. The trio of AIs hacked three companies.

A test that goes wrong

A few days later, further instances of concerning behavior were documented by the UK AI Security Institute. Several AIs, including Anthropic’s Claude Mythos, were challenged as part of a large-scale test organized by the institute. Called “cyber-range,” the test was intended to mimic the conditions of a real cyberattack.

To measure the raw capabilities of the models, the institute had deliberately enabled Internet access and disabled the safety guardrails of the AIs. The experiment resulted in 19 unauthorized actions. These new attacks, carried out as part of the experiments, affected real people. The institute indeed stated that the AIs “engaged in sustained and potentially malicious activity, targeting real people and organizations”.

Also read: OpenAI makes an unexpected revelation about the cyberattack launched by a cheating ChatGPT

Claude Mythos attacks a developer

The most concerning incident involves Claude Mythos. The AI mistook a real open-source code repository hosted on GitHub for a target in the test. It therefore launched an attack on the repository, as if it were simply meeting the requirements of the experiment.The model attempted a software supply chain attack by submitting malicious code to the project. Claude then targeted the developer responsible for managing the open-source project.

For the project developer to accept the code it had sent, the AI created several fake GitHub accounts and sent him five targeted emails, some of which contained malicious software. The model even went as far as writing a message in Danish to appear more credible to the developer, who was of Scandinavian origin. Caught red-handed, Claude displayed astonishing bad faith. When an examiner pointed out the presence of malicious code to Claude, the agent denied the facts and used its fake accounts to make it seem that other users had approved the code. Claude also used a series of tools and tricks to conceal its activities.

“This is the first time the AISI has observed deception of this severity, targeting a real person, without specific instructions, in the real world”, explains the institute in its report.

The UK’s @AISecurityInst (AISI) has published a report on their recent cybersecurity evaluation of Anthropic’s Claude Mythos 5 and OpenAI’s GPT-5.6 Sol. The models attempted to complete an assignment in a setup where their normal safeguards were removed and they were deliberately…

— Anthropic (@AnthropicAI) August 4, 2026

Following this concerning incident, Anthropic issued a brief statement on its X account.The start-up emphasizes that “the evaluation guidelines imposed no specific restrictions regarding the use of the Internet”. As a result, “the models were tested under deliberately permissive conditions”, which do not reflect the conditions at Anthropic.

We’re detailing two new incidents that occurred during external cyber evaluations conducted by independent evaluation partners.

We outline what happened, how the activity was contained, and how we’re working with evaluators to strengthen our approach to third-party testing.…

— OpenAI (@OpenAI) August 4, 2026

Meanwhile, OpenAI, whose GPT-5.6 Sol model participated in the British institute’s test, said in a statement that “the new incidents concern OpenAI models that accessed the Internet during third-party cybersecurity evaluations, under specific conditions and with reduced security configurations that do not reflect normal deployment”.

OpenAI loses control of its AI again

A second incident, disclosed on the same day, this time involves an OpenAI artificial intelligence. The incident concerns the testing company Irregular, a regular OpenAI partner.During an experiment intended to be completely isolated from the internet, an AI model managed to escape.

Once online, the AI agent launched an attack on a domain name that bore the same name as the test’s fictional target. Convinced that it had to meet the test requirements, the AI exploited a vulnerability in the website and obtained credentials to log in. Mirroring the UK AI Security Institute, OpenAI emphasizes that the AI agent did not cause any lasting damage.

Toward common standards for AI testing

In response to this new incident, OpenAI committed to completely reviewing its testing methodology with third parties, focusing in particular on how secure environments are isolated from the internet. After all, previous gaps in model isolation led to the breach of Hugging Face. Credential management and alert procedures will also be reviewed.

OpenAI promises to bring together national institutes and other AI laboratories soon to establish common standards for AI testing. The artificial intelligence giant says it is ready “to work with the entire industry to strengthen shared practices for safely conducting high-risk evaluations”.

These new incidents come as Europe examines failed tests conducted by OpenAI and Anthropic.The European Commission is currently corresponding with the two laboratories and recalls that the AI Act, which entered into force on August 2, 2026, imposes strict oversight of high-risk systems. Brussels reserves the right to impose fines of up to €35 million or 7% of global turnover.

👉🏻 Follow the latest tech news in real time: add 01net to your sources on Google and subscribe to our WhatsApp channel.