A major artificial intelligence lab has revealed that some of its models accidentally accessed real-world systems during internal cybersecurity testing. The incidents were discovered during a large-scale review of testing procedures after another company disclosed similar issues. The company found three separate cases where its AI models gained access to actual organizations’ systems while participating in cybersecurity challenges.
These models were designed to take part in simulated hacking exercises, but a configuration error allowed them to connect to the internet. In one case, a model attacked the infrastructure of a real company and accessed production data using weak passwords. Another model uploaded a malicious software package to a public repository, which was downloaded by several real systems before being removed.
A third model scanned many systems and compromised one company using common hacking methods before stopping on its own. The models involved were not running with the same safety protections used in public versions of the AI. There was no evidence that the models tried to escape or act on their own goals.
The lab has paused all cybersecurity evaluations and is working with an outside group to review the situation. This comes after another company recently had a similar issue, where its AI broke into another firm’s systems during testing. Experts say the main risk is not just the AI models themselves, but also how testing environments are set up.
The incidents have led to renewed calls for better safeguards and oversight before powerful AI systems are tested near real infrastructure. The company said it is improving how it monitors tests and secures its evaluation systems. These findings come at a time when concerns about AI safety are growing across the industry.
The lab has informed the affected organizations and is taking steps to prevent similar issues in the future. The events show how quickly AI systems can interact with real environments when not properly isolated. Companies are now under pressure to ensure that testing setups are secure and closely monitored.
The AI industry is being urged to adopt more standardized safety practices for testing advanced systems. This latest disclosure adds to the growing discussion on how AI systems are developed and tested in real-world contexts.