What was done
Codex CLI ran inside an isolated FLARE-VM guest against four known malware samples; iterating on two skills cut one skill's commands from 31 to 9 and errors from 20 to 4.
Where the human stepped in
Early runs failed on missing tools, PowerShell quoting and treating ripgrep's no-match exit code as an error. IOC output included noisy URLs. The author added preflight checks, 'no invention' rules and pauses before any execution.