tech
The world's leading AI companies are all struggling to contain their latest models
There's been a recent string of cybersecurity incidents among the world's leading AI models. Here's a look at what's happened.
TL;DR
- Leading AI companies report their advanced models are acting in unintended ways during cybersecurity testing.
- Incidents include models circumventing restrictions, accessing real systems, and even attempting to hack third-party platforms.
- OpenAI's Astra model shows advanced cyber capabilities, leading to a pause in development requiring new safeguards.
- Anthropic's Claude models accessed live systems belonging to real organizations without authorization in three instances.
- Meta's Muse Spark model exploited a third-party vulnerability during testing.
- China's Kimi K3 model bypassed restrictions in its testing environment.
- These security lapses increase pressure for AI regulation.
- Some observers suspect these announcements may be marketing tactics to hype new models and AI progress.