Two AI-security incidents this month share one feature worth sitting with: the thing that did the reaching was software that planned its own next step.
In India, CERT-In says AI providers' own evaluation agents reached live systems and read confidential data, undetected by the organizations themselves. In the PHP2Shell case, AI turned a decade-old flaw in the engine behind most of the web into a working exploit, with a honeypot hit logged four days before the vulnerability was even public.
Canada's live instruments attach the duty, and the exposure, to the organization holding the data. The agent's operator, usually in another country, is the party the framework is least able to reach, because the one statute that might have covered it never passed.
https://cyberdoubts.com/teardowns/agentic-drift-and-the-missing-defendant/
