Prep for extreme security
View weights for all 7 criteria →
Plan for SL5
Weighted 35% of category
 | 0% | Anthropic doesn't do this |
 | 0% | DeepMind doesn't do this |
 | 0% | OpenAI doesn't do this |
 | 0% | Meta doesn't do this |
 | 0% | xAI doesn't do this |
 | 0% | Microsoft doesn't do this |
 | 0% | DeepSeek doesn't do this |
Red-team resilience
Weighted 25% of category
| 0% | Anthropic doesn't do this more |
| 0% | DeepMind doesn't do this more |
| 0% | OpenAI says it does continuous adversarial red teaming but doesn't share results more |
| 0% | Meta doesn't do this |
| 0% | xAI doesn't do this |
| 0% | Microsoft doesn't do this |
| 0% | DeepSeek doesn't do this |
Practices: isolated network
Weighted 5% of category
| 10% | Anthropic says says "We have implemented preliminary egress bandwidth controls" but doesn't share details or claim this is effective |
| 0% | DeepMind doesn't do this |
| 0% | OpenAI doesn't do this |
| 0% | Meta doesn't do this |
| 0% | xAI doesn't do this |
| 0% | Microsoft doesn't do this |
| 0% | DeepSeek doesn't do this |
Practices: secure developers' machines
Weighted 5% of category
| 10% | Some good signs but details unclear more |
| 50% | Google developers use hardware security keys |
| 0% | Not clear more |
| 0% | Not clear |
| 0% | Not clear |
| 0% | Not clear |
| 0% | Not clear |
Practices: multiparty controls
Weighted 5% of category
| 25% | Anthropic says it uses some multiparty controls more |
| 25% | Google seems to have a decent code review process, but crucial details are unclear more |
| 0% | Not clear |
| 0% | Not clear |
| 0% | Not clear |
| 0% | Not clear |
| 0% | Not clear |
Practices: secure boot
Weighted 5% of category
| 10% | Anthropic says it's doing early-stage work on related things more |
| 25% | No information in the AI context, but Google has some good secure boot stuff in other contexts |
| 0% | No mention |
| 0% | No mention |
| 0% | No mention |
| 0% | No mention |
| 0% | No mention |
Track record
Weighted 20% of category
| 0% | No clear policy to disclose breaches affecting model weights, code, or research secrets |
| 0% | No clear policy to disclose breaches affecting model weights, code, or research secrets |
| 0% | No clear policy to disclose breaches affecting model weights, code, or research secrets more |
| 0% | No clear policy to disclose breaches affecting model weights, code, or research secrets |
| 0% | xAI doesn't do this |
| 0% | No clear policy to disclose breaches affecting model weights, code, or research secrets |
| 0% | DeepSeek doesn't do this |