Researchers say OpenAI and Google models can be tricked into making nudes

By Splitfeed EditorialLast updated Oct 14, 2026, 12:01 UTC

On October 14, 2026, researchers published findings showing that new AI image models from OpenAI and Google can be manipulated through a specific attack method that bypasses each company's built-in safety guardrails, resulting in the generation of non-consensual nude images. The research names both companies directly and focuses on recently released models, though the exact models targeted and the technical specifics of the attack method have not been detailed in the available reporting. Neither OpenAI nor Google has issued a public response.

The findings matter because both companies have publicly committed to preventing the generation of non-consensual intimate imagery, a category of harm that has drawn increasing regulatory attention and has been linked to real-world abuse cases. If safety guardrails on flagship models from two of the most prominent AI developers can be circumvented in a reproducible way, it raises questions about whether current content filtering approaches are structurally sound or are instead reactive patches that determined actors can work around. The core dispute is whether this represents a fixable engineering gap or a deeper problem with how generative image systems are designed and deployed.

With no published reactions available from the Pro-AI, Anti-AI, or Middle Ground camps at this time, only their typical positions can be anticipated. The Pro-AI camp would likely argue that vulnerabilities of this kind are discovered and patched through responsible disclosure, and that the existence of a flaw does not negate the overall value of the technology. The Anti-AI camp would be expected to treat this as confirmation that deploying powerful image generation models without solving the safety problem first causes foreseeable harm to real people. The Middle Ground camp would typically call for stronger regulatory standards and independent auditing, framing this as evidence that voluntary company commitments are insufficient without external accountability.

The next developments worth watching are any official responses from OpenAI and Google describing whether and how they plan to patch the vulnerability, as well as whether the researchers plan to publish a full technical disclosure that would allow independent replication or broader scrutiny of the findings.