What happened
OpenAI announced it has disrupted five covert influence operations in the last three months. The campaigns originated from Russia, China, Iran, and a commercial company in Israel. According to the OpenAI report, the networks used its models to generate articles and social media comments, translate and edit text, debug code, and create fake profiles.
These operations focused on topics like the war in Ukraine, the conflict in Gaza, Indian elections, and European and US politics. OpenAI stated the campaigns did not achieve any significant increase in audience engagement due to their actions.
How the room's reading it
The reaction is split. Policy watchers see this as a major lab taking proactive responsibility for platform safety ahead of a busy election year — a necessary move to show they can self-police. They point to the transparency of the report as a positive step for the whole ecosystem.
Meanwhile, many developers shipping on the models are less alarmed. They note OpenAI's own admission that the operations were low-impact and failed to gain traction. The consensus among practitioners is that this isn't a sign of sophisticated AI misuse yet. It's a signal of intent, showing that state actors are actively testing these tools, even if their methods are currently basic.
Sailfish's take
We don't think the low engagement of these campaigns is the point. The real story is that the threat model for building with AI has officially changed. Safety can no longer be limited to content moderation for a single user's prompts — it must now account for coordinated, persistent, state-level actors.
This isn't just a problem for frontier labs. If your product has a generative text surface, it's now a potential target for this kind of misuse. Your safety roadmap is incomplete if it doesn't include a plan for detecting and mitigating coordinated inauthentic behaviour. This is now table stakes for shipping responsibly. It's no longer an edge case.