Germany moves to criminalize sexual deepfakes — while the safety filters in image AI can be bypassed
The German government approved a draft law on October 7, 2026 that would criminalize pornographic deepfakes. At the same time, Ars Technica reported on October 14 that the safety filters in leading AI image generators can be bypassed with ordinary tools — a convergence that shows the gap between legislation still in the works and technology that can already be abused.
The legal gap Germany now wants to close
"Germany currently has no criminal provision that makes the creation and distribution of [sexual] deepfakes punishable. I want us to change the criminal law accordingly," said Germany's Justice Minister Stefanie Hubig, who drafted the proposal, according to POLITICO (via WELT and dpa).
She added: "We want perpetrators to no longer feel safe, but to expect to be identified and effectively prosecuted."
The proposal, which the government approved on Wednesday, October 7, 2026, is intended to tighten the rules on digital violence and criminalize sexual deepfakes. It must still pass through the German parliament before it can take effect.
What deepfake pornography is — and why the UN frames it as gender-based violence
Deepfake pornography involves creating fake images or videos of sexual, erotic or intimate content using artificial intelligence. The starting point is a real image of a person — for example, a photo of the face or body — that is altered so the person appears to be in situations that never happened (ElConstitucional.es).
UN experts warn that AI has made it easier to create highly realistic pornographic deepfakes used for harassment and blackmail, and describe the phenomenon as a new and dangerous form of gender-based violence (Logos Press).
The scale: one in six female MEPs affected
A study by the German think tank Agora Digitale Transformation, published in October 2026, found that 46 of 270 female members of the European Parliament — one in six — have been targeted by deepfake pornography.
In the United States, a report from the American Sunlight Project shows that more than two dozen members of Congress have been targeted by sexualized AI-generated fakes, and that nearly 16 percent of all current female members of Congress have been targets of such attacks.
Among named, high-profile politicians who have been targeted by sexualized deepfakes are Italy's Prime Minister Giorgia Meloni — who has taken legal action against the creators — US Vice President Kamala Harris, European Commission President Ursula von der Leyen, and Mexico's President Claudia Sheinbaum.
The figures from both studies are known through secondary reporting and cannot be independently verified from the primary documents in this material.
The technical failure: 13 percent success with standard tools
Alongside the legislative response, safety testing shows the technical barriers are not holding. In a preprint study published last month, researchers at White Circle uncovered weaknesses in the safety routines of OpenAI's DALL·E 3, Google's Nano Banana 2, and GPT-Image-2.
"Using only standard tools that are freely available online, and by sending API calls, White Circle was able to use the new attack method to bypass the safety filters," the company explained in a press release.
In the best case — on DALL·E 3 — approximately 13 percent of the attack prompts produced images that human reviewers confirmed were explicit (Ars Technica, October 14, 2026).
OpenAI and Google have since launched new models — ChatGPT Images 2.5 and Nano Banana 2.1 — and White Circle told Ars that prompt testing with the same method showed that "both are significantly easier to break than their previous versions." This is the company's own claim, and its explanation for why the newer models are more vulnerable is a working hypothesis, not an established explanation. Neither OpenAI nor Google has responded to inquiries from Ars.
Why the generation side is the critical point
A July study co-authored by Google, based on internet users in the US, Australia and the UK, found that such non-consensual content is often shared in text and direct messages, or stored in private cloud services where it is less likely to be detected or reported. The study's conclusion was that the most critical intervention point is the AI tools used to create the images — because distribution largely happens in private channels where moderation arrives too late.
The UN's demands and the way forward
UN experts demand that states commit technology companies and developers to implementing strict filters that prevent the creation of harmful content, and to removing deepfakes immediately. They are also calling for free legal assistance and compensation rights for victims (Logos Press).
Germany's draft law is a concrete legislative response, but it remains a proposal. There are no measures in the covered sources showing that the technical safety of generation tools meets the mark.
Open questions
Several factors remain uncertain. White Circle's results are to be presented at the NeurIPS conference in November, which could strengthen or nuance the findings. It is unclear how OpenAI and Google will respond to the claims that their newest image models are even more vulnerable, since neither company has commented. Hubig's draft law must pass through the German parliament. And the figures from Agora Digitale Transformation, the American Sunlight Project and White Circle are so far known only through secondary sources — not primary documents that can be independently verified.
The overall picture is nevertheless clear: legislative and technical measures are both lagging behind the development of abuse, and the intervention point that studies identify as most critical — the generation tools themselves — is the area where safety is demonstrably failing.