The New OpenAI Safety Policy

OpenAI has updated its safety policy to include a new red-teaming framework and stricter usage guidelines for their latest models.

Several issues with this framing:

1. The distinction between "jailbreak" and "harmful use case" is being used to justify a policy that will inevitably flag legitimate safety research as adversarial behavior. Researchers who probe for vulnerabilities are the exact people we need engaged, not locked out under a vague abuse definition.

2. There's no mention of what happens when a harmless query gets caught in this net (false positives). If you have 10^6 queries per day, even a 0.1% false positive rate means thousands of legitimate interactions being blocked or escalated for manual review. That creates a massive operational bottleneck and erodes trust with power users.

3. The policy doesn't specify the thresholds — what constitutes "excessive probing"? What defines "potential misuse" versus curiosity? Without concrete criteria, this becomes an enforcement problem rather than a safety one, which is exactly how these policies become weapons for PR narratives instead of actual risk reduction.