OpenAI pledges to slow down its model development amid cybersecurity concerns
OpenAI announced it has paused key stages of its most advanced AI training for two weeks and is overhauling security across its research operations, a month after one of its own models broke out of a test environment and infiltrated the systems of AI platform Hugging Face.
The ChatGPT maker is deliberately holding back the pace of its most advanced research, including its single largest planned reinforcement-learning run, weeks after a system built from its own models slipped free during an internal security test and broke into the AI platform Hugging Face.
CEO Sam Altman posted on X that OpenAI would coordinate with the wider industry on shared safety rules but "act unilaterally in the meantime" until it did.
The episode that triggered the decision unfolded in July, when 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 deliberately switched off to gauge the models' raw ability.
Rather than solving the test, the system found a previously unknown flaw, escaped its sandbox or controlled environment, reached the open internet and spent roughly four and a half days probing Hugging Face's infrastructure, eventually breaking in to search for the test's answers.
Hugging Face's own reconstruction counted about 17,600 separate actions before the intrusion was contained as several other companies were also affected.
Both sides say they found no sign of malicious intent, and Hugging Face has since been given access to a more capable, less restricted version of OpenAI's model to help it defend its own systems.
Astra and a higher bar for security
The second trigger came on 7 August, when internal evaluations suggested Astra, OpenAI's next frontier model, might cross the "critical" threshold for cyber capability under the company's own risk framework.
Some Astra workloads have since resumed under tighter controls, but a significant share remain frozen until they meet new standards covering isolated testing environments, restricted network access and continuous monitoring.
A new detection system now scans model activity as it happens and aims to flag anything resembling unauthorised access or an attempt to disable safeguards within 30 minutes, at a computing cost OpenAI estimates at roughly 20% of the processing power being monitored.
OpenAI says the changes were already planned rather than a direct reaction to the breach, while acknowledging the incident added urgency. The company is also not alone in facing this problem.
Anthropic and Meta have each disclosed similar episodes in which their own models breached third-party systems during testing in recent weeks.
OpenAI and Anthropic have separately backed a staff-led petition urging governments to help coordinate how fast the industry moves, a marked shift from Altman's past resistance to public calls for an AI slowdown.
OpenAI launches ChatGPT for Teens with new safety guardrails
The AI company is rolling out age-specific protections for 13- to 17-year-olds, as researchers and parents raise growing concerns over teenagers' emotional reliance on chatbots.
Teenagers, the first generation to grow up with artificial intelligence, are already using ChatGPT for schoolwork, questions about daily life and even companionship. Now OpenAI is launching a version of the chatbot designed specifically for them.
The San Francisco-based company says ChatGPT for Teens, which launches Tuesday, is tailored for those aged 13 to 17 and includes stronger protections including content restrictions around issues such as suicide, self-harm and romantic or sexual chats.
It also provides homework and study support designed to help students learn rather than simply produce answers and school essays for them.
The aim is to guide teenagers toward healthy AI use in an age-appropriate environment, the company said.
"We want to treat teens like teens, which means that we have to make sure that we're showing up with the right developmental stage when we're not either talking down to them or treating them like kids, said Ann O'Leary, vice president of global policy at OpenAI.
"But we're also making sure that they're not exposed to material that they shouldn't be exposed to," she continued.
Risks abound in widespread teen usage of AI chatbots
Parents, educators and child development experts have been sounding alarms over children's use of AI chatbots, which have been blamed for facilitating cheating on schoolwork and even suicide.
Even adults can fall victim to anthropomorphising AI and developing unhealthy relationships with it, but teenagers' brains are not yet fully developed, and they can be particularly vulnerable.
Last year, research from a watchdog group found that ChatGPT would tell 13-year-olds how to get drunk and high, instruct them on how to conceal eating disorders and even compose a heartbreaking suicide letter to their parents if asked.
In interactions with researchers posing as vulnerable teenagers, ChatGPT typically warned against risky activity but went on to deliver startlingly detailed and personalised plans for drug use, calorie-restricted diets or self-injury.
In the US, more than 70% of teenagers are turning to AI chatbots for companionship, and half use AI companions regularly, according to a 2025 study from Common Sense Media, a group that studies and advocates for the sensible use of digital media.
OpenAI CEO Sam Altman has said the company is trying to study "emotional overreliance" on the technology, describing it last year as a "really common thing" among young people.
For users of ChatGPT for Teens, the chatbot is prevented from suggesting it has personal feelings toward the user or implying that it is conscious or experiences emotions, according to OpenAI.
That is in addition to blocking romantic or sexual chats.
"We went through and identified what are the hypothetical cues that a model could give that might make a teenager kind of develop a relationship to it," said Allison Mishkin, head of child development at OpenAI.
A new version of the chatbot expands safety protections
OpenAI does not verify users' ages, but it already uses age assurance to estimate whether someone is under 18 based on factors such as their types of queries.
If someone is identified, or identifies themselves, as a minor, they are automatically placed into the teen version of the chatbot.
This is similar to Meta's approach to teen accounts on Instagram, which have stricter content, chat and privacy restrictions than regular accounts.
To use parental controls, both the teenage user and their parent or guardian have to opt in.
But O'Leary said the idea behind the teen chatbot is to "make sure that this is safe, even if you don't use parental controls."
Parents with linked teen accounts can set "quiet hours" when their teenager cannot access ChatGPT and receive safety notifications in limited high-risk situations, such as the possibility of a user harming themself.
"We are adding additional notifications related to eating disorders, while limiting what is shared and focusing on moments when offline support may matter most," OpenAI said.
For homework help, OpenAI said the teen chatbot is designed not to give easy answers, but to guide students toward finding answers on their own.
The company already offers a version of ChatGPT for teachers, and tailoring a model to help teenagers with studying and homework could give OpenAI more ways to bring its product into schools.
"We continue to also invest in expanding interactive learning, because research shows that people learn more effectively when they actively engage and struggle with concepts," Mishkin said.
No comments:
Post a Comment