OpenAI's AI hunters stopped 47 covert influence ops in 3 months

Three months. Forty-seven covert influence operations disrupted. OpenAI’s latest safety report doesn’t read like a typical corporate responsibility document — it reads like an intelligence briefing from a team that’s gotten very good at using AI as a defensive weapon. The June 2025 report details how the company’s expert investigative teams are wielding their own models as force multipliers to hunt down everything from cyber espionage campaigns to deceptive employment schemes.

What’s striking isn’t just the volume. It’s the range. The report catalogs abuse across social engineering, scams, spam, malicious cyber activity, and child exploitation. OpenAI frames this as a direct extension of their March submission to the Office of Science and Technology Policy’s U.S. AI Action Plan, where they argued for what they call “common-sense rules” aimed at preventing actual harms rather than hypothetical ones. The company is clearly drawing a line: enable AI through regulation that stops bad actors, not through regulation that chokes off innovation.

There’s a geopolitical dimension here that goes beyond standard trust-and-safety work. OpenAI explicitly names authoritarian regimes as a threat vector — governments using AI to amass power, control citizens, or coerce other states. That language is unusually blunt for a corporate publication. It signals that the company sees its defensive work as aligned with broader democratic interests, not just platform integrity. They’re effectively arguing that AI security is national security.

The report marks the second installment in what appears to be a quarterly transparency series. Three months ago, the numbers were different. Now they’re higher. The pace isn’t slowing down — and neither are the adversaries. OpenAI’s bet is that the same models enabling these attacks can be turned back on them faster than the attackers can adapt. Whether that calculus holds is the question hanging over every update like this one.