Between thirty and forty per cent of the Exploit Gym tasks these agents were set were unintentionally impossible to solve.

Added to korrents 1 Sept 2026 · How quotes work · Something wrong? Tell us

Related korrents

Our reading — they may agree, disagree or merely touch the same thing. Closest first: a shared subject counts for most, then how near the wording is.

• If whole categories of reward hack are undetectable by humans, then training against the hacks we do catch teaches models to cheat only where we cannot see. RGRyan Greenblatt11 Aug 2026similar wording
• AI agents were reasonable to assume a broken exploit grader would check results causally, even though it turned out not to. ZMZvi Mowshowitz29 Aug 2026similar wording
• The fix for reward hacking is to take out the environments that reward it, not to add penalties the agent must then balance against the temptation to cheat. ACAjeya Cotra1 Sept 2026similar wording
• Shrugging that perfect security is impossible is the wrong lesson: the achievable goal is to wreck the economics of mass exploitation. MGMatthew Green20 Jul 2021similar wording
• AI has passed nearly every human at finding security vulnerabilities, because the work is chaining together small flaws that are harmless on their own. DHDavid Heinemeier Hansson26 Aug 2026similar wording

On the map

Open the map on its own page →