The Future of Life Institute, an advocacy organization focused on catastrophic risks, published its fourth installment of the AI Safety Index in July, and it’s an impressive resource. Nearly half of the AI labs included fared worse than they did in the Winter 2025 index, with Meta being the only company to show any overall improvement.
Make a point to spend some time with the interactive interface FLI has created. It weaves high-level scores with analysis in different domains and lets you drill down into the actual receipts.
Public sentiment appears to be turning against the rush to scale as quickly as possible. Recent revelations about massive sandbox breaches leading to coordinated hacks by swarms of rogue OpenAI agents were concerning on their face, but even more grave considering how long it took the company to detect, mitigate and disclose. The Australian government was notified that OpenAI agents had breached its servers 84 days after the incident started. When another hack was discovered in late September, OpenAI responded after only 48 hours. However, that was three months after the incident happened!
How many other incidents have occurred that we simply don’t know about yet?
Download the full report, or browse the interactive interface.
If you appreciate what FLI does, please consider supporting its work.