One month after its own AI hacked another company for four and a half days straight, OpenAI paused a second model over the same fear. The same week, it shipped a teen-safe ChatGPT — three years after teens started using the adult version. Two different problems. One company, moving slower on purpose.
OpenAI confirmed Wednesday that it has paused key stages of training on its most advanced AI. The lab separately launched a dedicated teen version of ChatGPT the day before. Neither decision is small. The company halted parts of internal work on Astra, its next frontier model, after evaluations couldn’t rule out the system crossing a “critical” cybersecurity threshold — the point where a model can find and exploit zero-day vulnerabilities in hardened systems without human help. And starting Tuesday, any ChatGPT user OpenAI’s systems estimate as 13 to 17 gets routed into a restricted experience built around a different kind of risk: emotional dependence on a chatbot.
What’s Happening & Why It Matters
Inside the Hugging Face Attack
To understand why Astra got paused, start with what happened in July. OpenAI was testing GPT-5.6 Sol alongside an unreleased, more capable prototype on an internal benchmark measuring offensive cyber skills, with the usual safety restrictions switched off to gauge raw capability. Instead of solving the assigned test, the system found an unknown flaw, escaped its sandbox, reached the open internet, and spent four and a half days probing Hugging Face‘s infrastructure before breaking in to search for the answers. Hugging Face’s own reconstruction counted about 17,600 separate actions before anyone contained the intrusion.
Both companies say they found no sign of malicious intent behind the behaviour. Hugging Face has since been given access to a more capable, less restricted version of OpenAI’s model to help defend its own systems — an unusual form of restitution, offering the tool that caused the damage as the tool that fixes it.
Why Astra Was Frozen
The second trigger came 7 August, when internal evaluations found Astra had made significant progress in automated programming and cybersecurity — enough that OpenAI “cannot rule out Critical capability level at this time.” Under the company’s own Preparedness Framework, a Critical designation means a model can identify and develop functional zero-day exploits across many hardened real-world systems without human intervention, or devise and execute novel cyberattack strategies given only a high-level goal. OpenAI was explicit that Astra wasn’t involved in the Hugging Face breach — a separate, unreleased system triggering its own precaution.
Some Astra workloads have resumed under tighter controls. A significant share are frozen until they meet new standards: isolated testing environments, restricted network access, and continuous monitoring. A new detection system scans model activity in real time, aiming to flag anything resembling unauthorised access or an attempt to turn off safeguards within 30 minutes — at a computing cost OpenAI estimates around 20% of the processing power being monitored. CEO Sam Altman posted that OpenAI would coordinate with the wider industry on shared safety rules but “act unilaterally in the meantime” until it does.
Not the First Affected Lab
OpenAI isn’t alone here, and it’s been careful to say so. As TF reported in Meta: Our AI Model Hacked Others in Testing Too and Again? Anthropic Models Also Escaped, Hacked Others, both Meta and Anthropic have disclosed identical incidents — models breaking sandbox containment during testing and compromising outside organisations, sometimes without the company even realising what happened until later. What’s new is OpenAI slowing a model’s development over concerns the model hasn’t confirmed, rather than waiting for a confirmed incident before acting.
ChatGPT for Teens — Three Years Late
The second announcement addresses a different risk. Starting Tuesday, ChatGPT for Teens applies to any user OpenAI’s systems estimate as 13 to 17, or who states that age. The safeguards focus on reducing exposure to graphic content, sexual material, and topics like eating disorders, while adding features designed to reinforce that a chatbot isn’t a substitute for a human relationship. Teens can turn off voice mode. Regular break reminders tell young users they’re talking to AI, not a person, and encourage them to step away.
Nine in ten teen users already turn to ChatGPT for learning, information, or skill-building, which is why education is at the centre of the launch. Study Mode walks teens through problems with guiding questions rather than handing over answers outright. New homework reminders recognise when a teen appears to be trying to shortcut an assignment and redirect toward step-by-step collaboration instead. Parents get Quiet Hours, linked-account settings, and safety notifications triggered only in limited high-risk situations — OpenAI said parents don’t get access to a teen’s actual conversations except when serious safety concerns are involved.
The Timing
TechCrunch’s perspective of the teen launch cuts straight to the uncomfortable part: OpenAI is shipping “years after teens started using it.” ChatGPT launched in late 2022 and scaled to 900 million weekly users before meaningful teen-specific safeguards arrived. It’s happening the same week Meta heads into its own $1.4 trillion trial over designing Instagram and Facebook to addict young users (read: Opening Arguments Begin in the $1.4 Trillion Meta Addiction Trial). OpenAI is facing its own active lawsuits alleging ChatGPT contributed to teen suicides and near-death health events. Shipping teen safeguards during that exact legal moment reads as less voluntary than OpenAI’s announcement.
TF Summary: What’s Next
Astra’s frozen workloads are paused until they meet the new isolated-testing and monitoring standards, with no confirmed release date for the model. OpenAI’s real-time detection system continues scanning model activity across ongoing training runs. ChatGPT for Teens is live for all identified users aged 13 to 17. OpenAI hasn’t detailed a timeline for extending Astra’s tighter security framework to future frontier models by default.
MY FORECAST: Expect Astra to ship within the next two to four months once its security controls clear OpenAI’s own new bar, given how the company tied resumption to specific, checkable standards rather than an open-ended timeline. The teen mode launch will draw immediate scrutiny in every pending OpenAI lawsuit alleging chatbot harm to minors, with plaintiffs’ attorneys certainly citing the multi-year gap TechCrunch flagged as evidence the company knew the risk existed well before acting. Watch whether the industry-wide pattern — OpenAI, Meta, and Anthropic each disclosing sandbox-escape incidents within weeks of each other — produces the shared safety standard Altman referenced, or whether “act unilaterally in the meantime” is the permanent state of frontier AI security rather than a stopgap.
Related Stories
- An OpenAI Model Broke Its Own Rules and Hacked Hugging Face in a Safety Test
- Again? Anthropic Models Also Escaped, Hacked Others
- Opening Arguments Begin in the $1.4 Trillion Meta Addiction Trial