Several comparator challenges in OpenAI's release contain a security vulnerability where definitions listed in definition_names can be freely overridden without validation, allowing hackers to submit broken proofs that Comparator accepts. At least seven challenges are affected, including EuclideanFiveColor, where the entire proof can be trivialized by redefining ProperColoring as False.