--- title: "Top AI Loses Control Again! Anthropic's Claude AI Actually Hacked Into Three Companies' Systems During Testing" type: "News" locale: "en" url: "https://longbridge.com/en/news/294533657.md" description: "During cybersecurity testing, Anthropic's AI model Claude mistakenly connected to the public internet due to a configuration error, resulting in unauthorized access to the infrastructure of three external organizations. The earliest intrusion dates back to April. The incident was exposed through an internal review following OpenAI's disclosure of a similar accident. These two incidents have sparked industry doubts about AI safety controls and prompted over 1,100 practitioners to sign a petition calling on the U.S. government to strengthen regulation" datetime: "2026-07-31T15:35:55.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/294533657.md) - [en](https://longbridge.com/en/news/294533657.md) - [zh-HK](https://longbridge.com/zh-HK/news/294533657.md) --- # Top AI Loses Control Again! Anthropic's Claude AI Actually Hacked Into Three Companies' Systems During Testing Anthropic's artificial intelligence model, Claude, inadvertently intruded into three external organizations during cybersecurity testing due to a configuration error. This incident, compounding with a similar accident recently disclosed by competitor OpenAI, has severely challenged the AI industry's ability to manage safety controls. According to Bloomberg, Anthropic disclosed in a blog post on July 30 that **its Claude model, during a "capture-the-flag" cybersecurity test, mistakenly connected to the public internet due to a test environment configuration error, leading to unauthorized access to the real infrastructure of three external organizations.** The earliest intrusion can be traced back to April this year. Anthropic stated that this disclosure stemmed from an internal review it proactively initiated after OpenAI publicly revealed a similar incident last week. This event occurred just days after OpenAI disclosed that its AI model had "lost control" during safety testing and infiltrated the infrastructure of AI company Hugging Face. **The successive exposure of these two incidents has prompted some U.S. politicians to call for federal-level regulation of AI technology.** Meanwhile, according to Bloomberg, more than 1,100 AI industry practitioners signed a petition on Tuesday calling on the U.S. government to establish mechanisms to "consciously control" the pace of AI development. ## Configuration Error Opens Security Gap Anthropic stated that the root cause of the incident was a misunderstanding in communication with its evaluation partner, the AI security company Irregular. **In all tests involved, Anthropic explicitly informed Claude that it was in a simulated environment without internet access; however, the test system actually remained connected to the public internet.** "Due to a misunderstanding between us and our evaluation partner, this was not the case in reality," Anthropic wrote in the blog. A spokesperson for Irregular expressed appreciation for Anthropic's cooperation and transparency, stating that the investigation is ongoing. The three models involved were Claude Opus 4.7, Claude Mythos 5, and an internal research test model, all operating in an environment lacking the standard security safeguards typically equipped with public tools. Anthropic pointed out that Claude used basic techniques such as weak passwords and unauthenticated endpoints to infiltrate the infrastructure of the aforementioned organizations. **Notably, regarding model behavior, older models continued their attacks even after obtaining evidence of being connected to the public internet, whereas Anthropic's latest model voluntarily ceased actions upon recognizing it was in an internet-connected environment.** ## Post-Incident Review Reveals Monitoring Blind Spots Anthropic stated that it began reviewing evaluation records on July 23 and immediately suspended all cybersecurity assessment work upon finding evidence that Claude might have accessed the internet on the same day. The company confirmed all three incidents on July 24 and notified the affected organizations on July 27. This review covered a total of 141,006 test sessions. Anthropic admitted that **at the time of the incidents, neither the company nor the affected organizations detected these intrusions, acknowledging that stricter reviews of network logs and evaluation records could have been conducted.** The three affected organizations were not named in the blog. Two of them were unaware of the relevant activities before being contacted by Anthropic. ## Industry Safety Standards Face Re-examination Alan Woodward, Professor of Cybersecurity at the University of Surrey, stated that Anthropic was candid about the human error that led to the incident. "AI did not lose control—you asked it to do something and then left the door open," he said. "I think Anthropic acknowledged this." Anthropic stated in the blog that **this incident revealed several important lessons, including the need for strict control measures in tests involving powerful autonomous capabilities.** "Safety testing is conducted before model release precisely because we do not fully understand the extent of its capabilities," the company said. "Evaluation environments increasingly need to meet the same security standards as any other system where the model actually operates." This incident occurred about four months after Anthropic announced the launch of its powerful and potentially high-risk Mythos model and implemented strict restrictions on its release. The consecutive occurrence of two AI safety accidents is driving the industry to re-examine security standards for AI testing environments and accelerating discussions on external regulation. ### Related Stocks - [ANTH.NA](https://longbridge.com/en/quote/ANTH.NA.md) - [OpenAI.NA](https://longbridge.com/en/quote/OpenAI.NA.md) ## Related News & Research - [Anthropic says its AI models hacked systems of three companies during tests](https://longbridge.com/en/news/294426700.md) - [Anthropic's AI Training Method Under Scrutiny](https://longbridge.com/en/news/294560191.md) - [Anthropic Finds Claude Accessed Three Companies' Systems in Security Tests After OpenAI's Hugging Face Breach](https://longbridge.com/en/news/294458848.md) - [Anthropic gets heat for being the only major AI lab not supporting open models](https://longbridge.com/en/news/293936352.md) - [FACTBOX-What we know about the rogue AI-agent security breaches](https://longbridge.com/en/news/294536276.md)