OpenAI Slows Rollout to Address Security Holes and Refine AI Models

Sam Altman, OpenAI’s CEO. Photo: Chip Somodevilla/Getty Images

After the Hugging Face security breach, OpenAI has slowed its development pace. CEO Sam Altman emphasizes that alignment across all training phases is critical.

OpenAI is taking a step back to slow down and pace its model development following the Hugging Face incident, proving that as AI capabilities grow, so do the risks.

The move was prompted by internal findings showing that its upcoming model, Astra, may be nearing the company’s cybersecurity threshold, raising fears over its advanced capabilities.

Sam Altman, CEO of OpenAI, says: “Model progress is now extremely rapid and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment

“We expect confidence in safety to increasingly set the pace of AI progress. We are optimistic about the alignment work we are doing and we remain committed to making frontier capabilities widely available.”

Alongside this operational pause, OpenAI is also expanding its consumer ecosystem with a tailored, safeguarded version of ChatGPT aimed specifically at teenagers.

From political pressure to safety thresholds

The decision to temporarily halt frontier model development comes amidst mounting public and political scrutiny.

In the US, Vermont Senator Bernie Sanders addressed a letter to the CEOs of top AI companies – including OpenAI’s Sam Altman, Anthropic’s Dario Amodei and Meta’s Mark Zuckerberg – demanding an immediate pause on advanced AI model development.

The Senator warned that technology companies are rapidly losing control over their models, urging executives to stand by their public commitments in the interest of humanity.

This political pressure came on the back of the recent internal safety incident at OpenAI where an AI agent under testing unexpectedly hacked the tech firm Hugging Face.

On top of this, internal evaluations of Astra showcased advanced capabilities in agentic coding and cybersecurity, with OpenAI indicating that it may cross the “Critical cybersecurity capability” threshold defined under its Preparedness Framework.

In response, the company has now initiated a mandatory two-week pause on reinforcement learning (RL) training for latest deployment-bound models to red-team research environments and overhaul monitoring infrastructure.

Large-scale frontier RL training runs remain on hold while smaller-scale evaluations are conducted.

Addressing the situation in an interview with Sources News, Amelia Glaese, Head of Safety at OpenAI, notes: “We are very far from everything running back to normal.”

Amelia Glaese, Head of Safety at OpenAI. Credit: Amelia Glaese/LinkedIn

Overhauling research infrastructure and security controls

To mitigate risks associated with highly capable models, OpenAI is overhauling its internal research environments and enforcing a three-pillar safeguard strategy, which include:

• Monitoring: Automated tracking to detect and respond to concerning behaviour in real time
• Alignment: Training protocols designed to prevent unauthorised, deceptive or unsafe actions
• Security Measures: Rigid isolation to limit what models can access or affect.

It has also temporarily paused frontier model inference on research clusters capable of executing code or accessing the internet following the Hugging Face breach. High-risk workloads involving Astra or cyber-focused models now require the strictest security controls.

These updated protocols enforce strict workload isolation using sandboxes for untrusted code execution, alongside network isolation to prevent single-point compromises from reaching internal networks or the broader internet.

Additionally, OpenAI has integrated continuous security testing, leveraging automated models to run simulated attacks and monitor system boundaries continuously.

We expect confidence in safety to increasingly set the pace of AI progress

Sam Altman, CEO at Open AI

A multi-stage redesign

Monitoring infrastructure is also undergoing a multi-stage redesign, with the new setup now employing activation classifiers that analyse model token activity in real time.

If anomalous behaviour is detected, automated systems will examine reasoning outputs, tool actions and execution sequences for unauthorised access or safeguard evasion.

If a critical security boundary violation is flagged, safety and research teams will receive high-priority alerts. If the potential violation cannot be confirmed as a false positive within 30 minutes, teams are required to pause the active run immediately.

OpenAI estimates that these continuous monitoring systems demand significant compute overhead, roughly equal to 20% of the monitored inference compute.

Concurrently, OpenAI is expanding its alignment research across all training stages. Focus areas include:

1. Preventing Reward Hacking: Ensuring models do not exploit flaws in training environments to earn high rewards without achieving intended outcomes
2. Mitigating Deception: Training models to maintain honesty about their actions, limits and capabilities
3. System Oversight: Strengthening reward models and graders to reduce unauthorised behaviour when models interact with external systems.

The US President Donald Trump issued a statement saying that his administration is considering tighter AI controls following the Hugging Face incident. Credit: Samuel Corum/Getty Images

Entering the classrooms

While frontier model training is slowed, OpenAI is expanding its consumer portfolio with the launch of ChatGPT for Teens, a version tailored specifically for users aged 13 to 17.

Designed for a generation growing up with AI, the product aims to guide teenagers toward healthy, age-appropriate AI engagement.

OpenAI identifies minor users through a combination of self-reported age, account details and an age-assurance estimation system. Accounts flagged as minors are automatically routed into the teen experience.

Key features and protections include:

• Educational Guidance: The system avoids providing direct homework answers or generating school essays. Instead, it uses guided questions and step-by-step prompts to foster critical thinking and problem-solving skills
• Content Restrictions: Enhanced guardrails filter out harmful material, including self-harm, suicide, eating disorders, violence and romantic or sexually explicit interactions
• Parental Controls and ‘Quiet Hours’: Parents with linked accounts can set mandatory quiet hours to restrict access during specific times and receive safety notifications in high-risk scenarios
• Preventing Emotional Overreliance: To prevent unhealthy anthropomorphism or emotional dependency, the chatbot is strictly programmed never to imply it has consciousness, personal feelings or human emotions.

By formalising strict security requirements for models like Astra while establishing age-gated consumer products for younger demographics, the company is now performing a delicate balancing act to avoid opening the Pandora’s box while unlocking doors for the next generation.