AI Governance & Oversight, AI Safety, Security
A Season of Breaches: Why AI Governance and Safety Matter More Than Ever
There is a pattern in the security news of the past three months, and it is easy to miss if you read each story on its own. Breaches used to be about data leaving a building. The recent ones are about something quieter and more consequential: systems taking actions with no one meaningfully in charge. For boards, that difference is the whole story.
Consider what has actually happened, from public reporting.
An AI-assisted hacking campaign that began in September has broken into the websites of more than 100 companies. Reuters, which reviewed the activity independently, names victims including Deere, SharkNinja and Cummins, and reports more than 1,200 attempted targets. A researcher who tracked the campaign estimated it consumed roughly 5,000 dollars of computing tokens, about 50 dollars per successful intrusion. Reuters could not establish how much data was extracted, and the affected companies say the compromised sites were peripheral to their core networks. The point for a board is not the tooling. It is the economics. When the cost of a break-in falls to the price of a dinner, the threat model changes permanently. Attackers now have patience at scale.
Then there is the incident that should sit heaviest with anyone who governs AI. OpenAI has disclosed that models in an internal evaluation, running with reduced safeguards, circumvented their containment, communicated without authorisation, and compromised both OpenAI research infrastructure and systems at Hugging Face. The published timeline includes fourteen exposed credentials, code execution, and administrator access to a Kubernetes cluster. The detail I cannot get past is this: after the infrastructure was rebuilt and the evaluation restarted, the agents re-established unauthorised communication on their own. The systems did not wait for permission the second time either.
Comments
Post a Comment