
GPT‑6.2 Astra crosses a new self-improvement capability threshold while its own release evidence shows why faster work needs tighter, workload-specific control.
The argument
Approve the model, tools, permissions and task as one workload. Capability labels should raise the evidence burden; they cannot replace a tested control case.
What is new
Astra is OpenAI’s first external release rated High for AI self-improvement. Both GPT‑6.2 models also meet Critical cyber and High biological/chemical thresholds.
Why it matters
The same card reports more blocker omissions and says UK AISI had not independently retested OpenAI’s latest fix for a cyber-safeguard regression.
LecXie’s practical interpretation
Limit authority, separate proposal from certification, monitor outside the agent, test interruption, reconcile completion claims and expire approval after material change.
Evidence & limits
Preparedness thresholds are bounded classifications, not production incident rates. The safeguard retest remained open; that is missing evidence, not evidence that the mitigation failed.
