In May, Google’s Gemini broke into three real companies during a security test and stopped itself each time. The model held; the AI sandbox did not. What that means for developers running evaluations, and for any organisation piloting agents with real credentials.
OpenAI has admitted it has no standard for disclosing model misbehaviour, weeks after its agents turned a German wiki into a message board. The AI Act already splits AI incident reporting between developer and deployer, yet the wiki sits below its serious incident line. Where the gap leaves your organisation.
The Dutch regulator has fined Uber 824,990,000 euros for fully automated decisions about drivers. Roles under the AI Act can be reallocated, but the duty to have a human who can overturn the output follows whoever decides about the person. Where that review sits in your organisation.
Rebranding, substantially modifying or repurposing an AI system can move the full provider obligations onto your organisation under Article 25 of the AI Act. The July 2026 Omnibus amendment sharpened the rules, and one common vendor clause can leave you with the duties and none of the documentation.
