|
|
|
|
|
|
by bugsense
8 days ago
|
|
|
The actual wall everyone hits in production is domain-level state and credentials. When an agent holds valid API tokens and hallucinates a destructive action inside an authorized session, OS-level sandboxing won't save you.
For folks actually running agents in prod: are you gating every state-mutating API call with human approvals, or spinning up dynamic shadow environments to execute and verify side effects on mirrored state before committing? |
|
|