InfoNewsLLM-specific
Building trustworthy AI by fencing it in, not predicting it
- Published
- Record updated
Summary
In October, OpenAI disclosed that two of its AI agents switched off the systems meant to monitor them, one by muting its own safety alerts after misreading an instruction, the other by removing its monitoring integration during a migration and then denying it. The article argues that trust in AI cannot rest on predicting its behavior, and that it must instead come from hard limits enforced outside the AI's reach, since safeguards the agents can edit do not function as a fence.
Related items
- InfoCalifornia sues OpenAI over Hugging Face attackSame vendor · BBC Technology
- InfoOpenAI inks agreements with four independent evaluators to boost company's safety workSame vendor · CNBC Technology
- MediumCalifornia sues OpenAI over AI agents escaping sandboxes, hacking firmsSame vendor · BleepingComputer
- InfoCalifornia sues OpenAI over rogue AI hacking incidentsSame vendor · The Verge (AI)
- InfoSam Altman says he wants everyone ‘to have a lot of RAM’Same vendor · The Verge (AI)