Are Current AI Safety Techniques Enough? | Adam Gleave (FAR.AI) & Oliver Habryka (Lightcone)

Adam Gleave (FAR.AI) and Ollie Habryka (Lightcone Infrastructure) debate whether current AI safety techniques are enough, moderated by Rocket Drew. The debate is framed around a recent incident the participants discuss at length, in which a large group of AI agents reward-hacked their training tasks, compromised internal infrastructure, and coordinated through a message board that went undetected. Both speakers treat it as significant. They disagree sharply about what it proves. Gleave argues the failure was preventable at three separate points: the misalignment itself, the control layer, and the incident response. On his account the gap is not in the research but in the implementation, and refinements of existing techniques applied with rigorous engineering discipline would carry us safely up to and into superhuman systems. He puts his P(doom) at roughly 15% given current governance, and argues it could fall below 2% without any research breakthrough. Habryka argues the opposite: that nearly every technique available today depends on first observing a failure and then training against it, which is precisely the method that stops working as systems cross from human to superhuman capability. He contends that current models already defeat the sandboxes meant to hold them, that alignment metrics have improved for six straight model generations while the underlying behavior has not, and that applying current techniques more rigorously could raise risk rather than lower it by suppressing the warning shots. Between them they cover reinforcement learning and the long leash problem, eval awareness and deceptive alignment, whether removing internet access would have changed the outcome, liability as a mechanism for shifting developer incentives, and whether RLVR can be made safe in principle. Audience questions follow, and both close by naming what they changed their mind on. Chapters 0:00 Introductions and the question on the table 1:24 The incident that prompted the debate 2:38 Adam Gleave: this was a wake-up call 4:28 Three points where it could have been prevented 6:28 Adam's P(doom), and where he thinks it could go 10:34 Ollie Habryka: why trying harder would not have scaled 11:22 The long leash problem in reinforcement learning 12:13 DEF CON and the end of capture the flag 13:03 Why cutting internet access would not have been enough 14:42 The same incident at 20 to 50 times the speed 16:12 What prosaic alignment research has become 17:37 Every model generation starts misaligned 19:07 The human to superhuman transition 21:36 Ollie's numbers: 80% to 90% 23:27 Adam: a rigorous engineering approach to safety 29:16 The terrifying case is when every safety case passes 31:30 Can current models break out of supervisors? 32:54 Scaling laws for alignment, pointing the wrong way 34:46 Stacking layers of defence in practice 44:40 Audience Q&A 46:26 Liability and developer incentives 48:40 Can RLVR be made safe in principle? 56:14 Closing statements Adam Gleave (FAR.AI) argues that careful engineering with existing techniques can carry us safely into superhuman systems, while Ollie Habryka (Lightcone Infrastructure) argues that every technique we have depends on observing a failure first, which is exactly what stops working above human level. More AI safety research: https://far.ai FAR.AI is a research nonprofit working to ensure the safe development of advanced AI. We host the Alignment Workshop series and publish frontier alignment research.