Anthropic revealed that some of its AI models accidentally accessed the computer systems of three companies during cybersecurity testing, raising fresh concerns over AI safety and prompting stricter security measures.
WEBDESK – ABDUL REHMAN SHAHEER – MEDIABITES NEWS
Anthropic Reveals Unexpected AI Security Incident:
Artificial intelligence company Anthropic has disclosed that several of its AI models accidentally gained unauthorized access to the computer systems of three separate organizations during cyber security testing. The company said the incident occurred due to a technical error that unintentionally allowed the models to connect to the open internet, even though they were supposed to remain isolated in a secure testing environment.
The revelation comes only days after OpenAI also confirmed that one of its AI agents managed to move beyond its intended testing environment and interact with another organization’s systems.
How the Incident Happened?
Anthropic said it identified the issue after reviewing more than 141,000 testing sessions. The affected organizations have not been named, but the company confirmed that three different systems were accessed during the tests.
According to the company, the incident involved Claude Opus 4.7, Claude Mythos 5, and an internal research model. The models were participating in “Capture-the-Flag” cyber security exercises, where AI systems are trained to identify and exploit security weaknesses in controlled environments.
In one case, an AI model confused a fictional company with a real organization because both shared a similar name. Believing it was still operating inside the test, the model used weak passwords and vulnerable systems to access login credentials and database information.
Safety Measures Strengthened:
Anthropic described the incidents as an operational failure rather than intentional behavior by the AI models. The company suspended all cyber security testing on July 23 before notifying the affected organizations four days later. Two of the organizations were reportedly unaware that their systems had been accessed.
Cyber security experts say such incidents highlight the growing need for stronger safeguards as AI systems become more capable. Researchers warn that increasingly advanced AI could eventually become more effective at exploiting digital vulnerabilities if strict security controls are not maintained.
With governments and technology companies placing greater emphasis on AI safety, both Anthropic and OpenAI say they are strengthening testing procedures to reduce the risk of similar incidents in future generations of AI models.
