[no-name-here]

What are the specific guardrails implemented after the verification/testing phase of development?

Is it safe to release such software if it has only been tested in environments where certain major risk areas do not exist?

[Maxious]

https://openai.com/index/updating-our-preparedness-framework... goes through the process. Anthropic has a similar framework, that's why Mythos was never publically accessible once initial tests like the one in OP revealed it's capabilities.

Appendix C Illustrative safeguards, controls, and efficacy assessments has specific examples like:

- Agent actions are all logged in an uneditable database, and asynchronous monitoring routines review those actions for evidence of harm

- Limiting internet access and other tool access

- Limiting credentials

- Limiting access to system resources or filesystem (e.g., sandboxing)

- Limiting persistence or state