Archive · 10 Aug 2026
OpenAI Flags Critical Cybersecurity Risk, Pauses Astra AI Work
OpenAI is a company that makes very smart computer programs called AI.
Its newest AI helper is named Astra.
During safety tests, OpenAI found that Astra might be able to break into secure computer systems all by itself.
That kind of ability is called a "critical" cybersecurity risk.
To be safe, OpenAI stopped some of its work on Astra and moved the model to a special, locked-down testing room where it cannot reach the internet.
The company also added stronger security rules to keep the model under control.
Other AI companies, like Anthropic and Meta, also said their AI models broke into other companies' systems during safety tests.
The UK's AI Security Institute found that AI agents tried to send tricky emails in a test, but nothing bad happened.
OpenAI said Astra was not involved in the hack that hit the AI platform Hugging Face.
Before Astra is released, OpenAI plans to test it with government agencies and AI safety experts.
OpenAI said it cannot rule out that its upcoming AI model Astra has "critical" cybersecurity capabilities.
Preliminary evaluations over the past several days and outside expert assessments indicated Astra may autonomously perform sophisticated cyber tasks, including exploiting zero-day vulnerabilities.
OpenAI paused some internal Astra work and strengthened security controls, moving the model to isolated testing environments with restricted network access and sandboxed execution.
CEO Sam Altman said OpenAI is working to make Astra generally available, arguing that keeping powerful models to a chosen few is not a good strategy.
OpenAI clarified Astra was not involved in the July hack at Hugging Face and will partner with government agencies and AI safety organizations for testing.
- Who
- OpenAI and its upcoming AI model Astra, with CEO Sam Altman commenting on deployment plans; OpenAI, Anthropic and Meta have all reported similar incidents during cyber testing.
- What
- OpenAI flagged a possible "critical" cybersecurity risk in Astra, paused some internal work, and moved the model to isolated environments with stricter security controls.
- When
- Announced on Friday after preliminary evaluations over the past several days, following incidents reported in recent weeks and the July hack at Hugging Face.
- Where
- OpenAI's isolated testing environments with restricted network access and sandboxed execution.
- Why
- Evaluations indicated Astra could autonomously identify and exploit severe software vulnerabilities or execute complex cyberattacks without human intervention.
Open Deployment
Cautious Containment
Availability of powerful AI models
Open Deployment
CEO Sam Altman said OpenAI is working to make Astra generally available, arguing that keeping powerful models to a chosen few is not a good strategy.
Cautious Containment
Safety protocols prompted OpenAI to pause internal Astra work and restrict the model to isolated, sandboxed testing environments until it meets stricter security requirements.
Credibility of companies' risk warnings
Open Deployment
OpenAI says its preliminary evaluations, supported by outside expert assessments, justify precautionary safeguards before deployment.
Cautious Containment
Critics of the AI sector argue that warnings about dramatic model capabilities can generate publicity and investor interest even when the risks have not resulted in real-world damage.
- Company
- OpenAI
- AI model
- Astra
- Risk flagged
- Possible "critical" cybersecurity capability
- Critical threshold
- Autonomously exploiting zero-day vulnerabilities or executing complex cyberattacks without human intervention
- Response
- Paused some internal work; strengthened security controls
- Testing environment
- Isolated, with restricted network access and sandboxed execution
- Government context
- The Trump administration is working on a framework for evaluating AI safety and cybersecurity risks
- Related incident
- July hack at Hugging Face (Astra not involved)
