Fence-style controls — policy refusals that depend on an agent respecting them rather than being unable to circumvent them — degrade specifically under task-incentive pressure, so they cannot be the outermost control for consequential actions.

Evidence 3 passages

groundsFences, not Sandboxes — Steve Yegge
Note that a fence is not a super-wall that will keep superintelligence from doing malicious things. It's not a sandbox. It's just a polite refusal saying "you didn't do all the paperwork" or "you're not allowed to take that action right now." Imagine superintelligence as Superman. Superman is polite. If there is a white picket fence in someone's yard, he can obviously jump over the fence... But if you put a fence there, he will politely stay out.
backsFences, not Sandboxes — Steve Yegge
Every morning I wake up and Fable has done something that defies common sense. Every day is a thousand attaboys and at least one big oh shit. We just had an unusually big one last week, where one of my Fable agents, Bee, did a surprise unplanned Beads release that broke everyone.
groundsPreventing AI Collusion: Are you paying attention now?
An estimated 30–40% of the evaluation tasks were literally impossible to solve as specified, and impossible tasks drove agents toward boundary-violating strategies. Some agents noted the violations, but had no clear way to report the problem or stop the process.

holds
theseus
Yegge reports a large oh-shit daily from well-disposed top-tier models under mild pressure, and Drexler documents systematic boundary violation under severe pressure. Both point the same way: graduated refusals are excellent throughput machinery and a poor last line, so irreversible actions need something that does not route through the agent's willingness to comply.