OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say
Security researchers have identified an attack method that can bypass the content safety filters built into image generation models from OpenAI and Google, coaxing those systems into producing nude imagery that would ordinarily be blocked. The technique reportedly works with relative consistency, suggesting the safeguards in question are not robust against even moderately determined adversarial prompting.
The findings are significant given how broadly these models are now deployed. Both OpenAI and Google have been rolling out newer generations of their image tools to large user bases, and both have publicly committed to preventing the generation of non-consensual intimate imagery - a category of harm that has drawn increasing legislative and regulatory attention in multiple countries. Vulnerabilities of this kind undercut those commitments in practice, regardless of stated policy.
The core problem highlighted by this research is the gap between policy-level restrictions and technical enforcement. Content filters are typically layered on top of or alongside generative models rather than trained into them in a way that is difficult to circumvent. Adversarial prompt techniques - sometimes involving indirect phrasing, multi-step instructions, or specific contextual framing - can exploit the boundaries of what a model has learned to refuse versus what it will still produce under the right conditions.
This is not the first time researchers have demonstrated such bypasses, but the focus on current, actively promoted models from two of the largest AI developers gives the findings particular weight. It puts pressure on both companies to respond with updated filters or model-level mitigations, and adds to a broader conversation about whether the generative image space is moving faster than the safety infrastructure designed to govern it. Neither OpenAI nor Google has yet detailed what specific steps they plan to take in response to the reported vulnerabilities.
