Skip to content- Sources: primary
- Summary: The video describes a four-stage pipeline: a scout agent narrowed kernel master to about 5,000 code targets reachable by unprivileged users, an analyst produced candidate bugs for about $316,000 in tokens, a triage agent reproduced them in VMs with the kernel memory checker, and exploit agents built working root exploits with KASLR and SMEP enabled. The run cost about $1,213,000, or about $1,280 per exploit, compared with 166 local privilege escalations at about $3,100 each from a GPT-5.6 Sol run in July. OpenAI withheld technical details and says an unpatched low-privilege account should be treated as equivalent to root.
- Why it matters: Shared hosts, CI runners and agent sandboxes that rely on kernel privilege boundaries need prompt kernel patching and a stronger isolation layer such as a VM.