Correction: When unmonitored, it nearly always *refused [to produce harmful content]. But when monitored, it faked alignment 12% of the time.

→ View original post on X — @anthropicai