It seems like OpenAI are trying to frame this as if AI did all the work here, but it's not really true given what Calif say (https://calif.io/php2shell/):
> How did we find this?
> No special harness, or access to models that weren't generally available to everyone were used in the initial bug discovery. I focused on something I had a background in, knew its vulnerability history well, and held a theory that specific bugs lurk within the codebase. This shows that "know your target" is still as applicable as ever, even in the LLM dominated VR age. No model is simply capable of unearthing "all bugs" with simple prompts. At least for now.
> There are some practical research implications here. LLMs may have ingested public historical vuln data, but in their default variant analysis tends to move from the newest and most prominent bugs toward nearby older code. They do not reliably begin with obscure twenty-year-old exploitation primitives. Keeping your own mental catalog of a target's vulnerability history lets you prompt that direction explicitly.
> How did we find this?
> No special harness, or access to models that weren't generally available to everyone were used in the initial bug discovery. I focused on something I had a background in, knew its vulnerability history well, and held a theory that specific bugs lurk within the codebase. This shows that "know your target" is still as applicable as ever, even in the LLM dominated VR age. No model is simply capable of unearthing "all bugs" with simple prompts. At least for now.
> There are some practical research implications here. LLMs may have ingested public historical vuln data, but in their default variant analysis tends to move from the newest and most prominent bugs toward nearby older code. They do not reliably begin with obscure twenty-year-old exploitation primitives. Keeping your own mental catalog of a target's vulnerability history lets you prompt that direction explicitly.