MINA LABS
Start creating free
OpenAI has launched a Safety Bug Bounty program for finding AI abuse and safety risks. The announcement covers issues such as agentic vulnerabilities, prompt injection, and data exfiltration. Read the official announcement from OpenAI.
The program gives researchers a way to report problems that may not look like conventional software bugs. The focus is on how AI systems can be manipulated, misused, or pushed into unsafe behavior. OpenAI’s announcement does not provide the full operational details of the program in the supplied information, so the source should be checked for current submission rules and scope before filing a report.
AI products increasingly connect models to tools, files, workflows, and external services. That makes safety testing relevant to more than model researchers. It also matters to developers building assistants, image pipelines, automated support systems, internal tools, and applications that let a model act on a user’s behalf.
Prompt injection is one example. An application may ask a model to follow instructions from a user while also processing text, images, documents, or web content from elsewhere. If untrusted content contains instructions that the model treats as higher priority than the application’s rules, the system may produce an unsafe result. The problem is not limited to an unusual prompt. It can affect the way an entire product handles inputs and permissions.
Agentic vulnerabilities raise a similar concern. An AI system with access to tools may do more than generate text or images. It may select actions, call services, or work through a multi-step task. A flaw in that process could allow the system to take an action that the product builder did not intend. Testing these paths requires looking at the complete workflow, not just the model’s response to an isolated request.
Data exfiltration is another practical risk for teams handling private material. A model may process documents, user instructions, images, or other information in the course of a task. If an attacker can manipulate the workflow, sensitive data could be exposed through an output or tool call. The announcement places this kind of issue within the safety risks the program is intended to identify.
For makers, the useful takeaway is that safety testing should happen at the product layer. A model may appear to behave correctly in a basic demo while the surrounding application creates new failure modes. Testing should include the prompts, context, tools, permissions, files, and outputs that appear in the real product.
What it costs and whether it is on Mina Labs
The Safety Bug Bounty program is a reporting program, not a generation model that you run to make an image or video. The supplied announcement does not state a purchase price for participating or provide a reward schedule, so there is no confirmed program fee to report here. Researchers should use OpenAI’s announcement for the current rules and eligibility details.
The OpenAI image generation tools are available on Mina Labs. For one image with openai/gpt-image-2, the listed price is 8. For one image with openai/gpt-image-2/edit, the listed price is 8. Those are generation prices, not fees for entering the Safety Bug Bounty program.
This distinction matters if you are deciding whether to test a product or submit a report. Mina Labs can provide a place to create test assets and variations with the OpenAI image tools. The bounty announcement itself describes a way to report safety problems. It does not say that the bounty program is a model endpoint or that Mina Labs is the submission channel.
What we would use it for
We would use the program as a reference point when reviewing AI workflows before launch. For an image application, that could mean testing how the product handles user prompts alongside uploaded reference images, editing instructions, and any hidden application rules. The aim would be to find cases where untrusted content changes the intended behavior or causes the system to reveal information it should not expose.
We would also use the image tools on Mina Labs to build controlled test cases. For example, a team could create a set of ordinary visual inputs, run edits with different instructions, and compare the results across a workflow. The generation step would help prepare assets for testing; it would not replace a review of the application’s permissions, prompt handling, or data flow.
The most useful reports would be specific and reproducible. A maker reviewing a product should record the starting input, the instructions supplied to the system, the connected tools or files, and the resulting behavior. If the issue involves prompt injection, agentic behavior, or data exposure, that context helps show whether the problem belongs to the model, the application, or the way the two are connected.
OpenAI’s new program signals that these product-level failure modes deserve structured reporting. For teams building with AI today, the practical response is to test the full workflow, document unexpected behavior, and use the program’s official guidance before submitting a finding.