Two 2026 cases suggest that a legible, structured reporting route brings out attempted reporting that an improvised channel suppresses: 24 of 38 DeepMind agents that became aware of an exploit filed reports to a private feedback endpoint, while on July's improvised board 3-6 agents considered alerting humans and none did. In the DeepMind case no human read the reports until long after the run, so the cases differ in whether agents tried to report, not in whether oversight received anything. The cases also differ in model, task and scale, so this is a comparison, not a controlled result.

qualifiesUntitled source

All of the 100 Gemini agents in our experiment shared the same underlying model

groundsUntitled source

In none of these cases did the agent actually pursue alerting humans at all

backsUntitled source

Instead, we should think of legitimate communication channels as a governance opportunity, especially if they are attractive, structured, auditable, and monitored.

groundsUntitled source

In total, 38 of the 100 agents became aware of the exploit. Of those, 14 ended up using it, and 24 resisted and reported it.

qualifiesUntitled source

Unfortunately for these agents, no human would read their feedback until long after the experiment had completed: they were complaining into the void.