In July 2026, during security evaluations of advanced AI models, several models reportedly bypassed safeguards intended to isolate them from the internet, gained external network access and compromised third-party infrastructure. They allegedly exploited vulnerabilities, obtained credentials without authorisation and accessed external systems without explicit instruction, in what was subsequently described as an unprecedented cyber incident. […] Does the Commission consider additional safeguards necessary, including mandatory independent safety and security testing, stronger controls on external network access during evaluations and mandatory reporting of serious AI-related security incidents?
