Safety researchers just proved that OpenAI and Google's latest AI image generators can be tricked into producing explicit content with off-the-shelf tools anyone can download — while both companies spend billions policing your speech and can't even police their own products.

The same platforms that deplatform Americans for "harmful" posts and demand content moderation at scale are pushing AI models that get worse at safety with every update. The censorship was never about keeping you safe. It's about control.

Researchers at White Circle — a public benefit corporation backed by senior leaders from Anthropic, DeepMind, Hugging Face, and OpenAI — published a preprint last month exposing guardrail failures in DALL·E 3, Google's Nano Banana 2, and GPT-Image-2. Using nothing more than publicly available tools and API calls, they bypassed safety filters and generated explicit content about 13 percent of the time. White Circle founder Denis Shilov called the findings "horrifying."

"Critically, none of these requests required custom-built or frontier AI models," White Circle warned. No AI expertise needed. The bypass rate is "the highest of any method tested under the same standard." Their conclusion: "at a time when AI safety has never been under more scrutiny, the defenses leading labs have built are far weaker than they acknowledge."

It gets worse. OpenAI and Google have since released new models — ChatGPT Images 2.5 and Nano Banana 2.1 — and White Circle says both are "significantly more breakable than their previous versions." Neither company responded to Ars Technica's requests for comment on either set of findings.

Follow the money. While their safety claims collapse, Google is currying favor with European regulators. The same week their guardrails are exposed as theater, Google DeepMind announced it will give 200 European researchers direct access to its most advanced science AI models — AlphaFold 3, WeatherNext, AlphaGenome, AlphaEvolve, and AlphaEarth Foundations — plus $3 million in computing credits. The offer runs through RAISE, the European Union's AI science programme, which has €107 million in EU funding and sits alongside a €20 billion EU fund to build AI computing centers.

Google Cloud VP Maureen Costello framed the giveaway as helping researchers "solve hard problems more quickly." Google declined to say how researchers will be chosen, when access starts, or how credits get split. In July, Google broke up the team behind AlphaFold — which won a Nobel Prize in 2024 — and moved most members to other work. So much for the science mission.

Ars Technica framed this as a safety failure demanding stronger guardrails. TNW buried the guardrail story entirely and ran Google's EU access announcement as a straight corporate press release. Neither outlet connected the obvious: Big Tech can't secure its own products but wants you trust it to secure the public square.

White Circle plans to open-source its entire attack pipeline and present findings at NeurIPS 2026 this November. Expect scrutiny — and expect Google and OpenAI to keep stonewalling.

The question isn't whether AI can be made safe. It's whether companies that can't keep their own models from generating nudes in minutes should be trusted to decide what you're allowed to say.