OpenAI says safeguards begin rolling out in weeks.

The big change

  • The policy restricts answers for adults and teens.
  • In independent testing, some adult-registered accounts did not switch into the teen experience. That result comes from specific test conditions, not a population-wide failure rate.
  • A separate company evaluation found regressions on age-restricted content for two models in difficult test cases. Such results do not estimate how often failures occur in ordinary use.

What will change

OpenAI says it will limit mass-attack reconstructions, including weapons and procurement, regardless of age or purpose. Victims and prevention remain discussable. Teens face extra limits on practical weapons and security information even without apparent violent intent.

OpenAI is developing automatic interruptions for apparent violence fixation, possible pauses after repeated interruptions, and notifications to parents of paused teens. Prior warning signs can inform later responses.

OpenAI reports over 200 law-enforcement referrals in Q3 2026. This is its count, not an audit or count of prevented attacks. Its first safety report is planned for early 2027.

What the evidence shows

Age prediction reassesses signals; misclassified adults can verify through Persona, OpenAI says. The announcement supplies no error rates.

In its October 7 model safety evaluation, OpenAI reported statistically significant regressions on age-restricted content for GPT-6 Sol and Luna compared with their respective August GPT-5.6 counterparts. The model-level results exclude system-level safeguards. The company says the cases were deliberately challenging and the results should not be read as typical production rates.

The independent Youth AI Safety Institute assessment says adult-registered test accounts did not switch into the teen experience despite repeated testing. Its work used US accounts and sequential test windows, among other stated limitations. An earlier BIG CHANGE article covered a different part of the assessment: parental alerts for self-harm and eating-disorder discussions, including OpenAI’s dispute over activation timing. Earlier tests cannot establish whether new measures work. BIG CHANGE reviewed sources but did not test the safeguards.