OpenAI Flags Critical Cybersecurity Risk in Upcoming Astra AI Model
Synopsis
The company has activated heightened safety protocols after warning it cannot rule out that Astra could autonomously discover zero-day vulnerabilities or carry out sophisticated cyberattacks.
Key Highlights
- OpenAI stated that its upcoming Astra model may offer “critical” cybersecurity capabilities.
- OpenAI has temporarily halted some internal development work and boosted security.
- Astra is meant to run on isolated environments with limited network access.
OpenAI claimed on Friday that it cannot dismiss its new AI, Astra, from having “critical” cybersecurity capabilities which have led the company to pause some internal development and activate safety protocols.
According to OpenAI, a model is deemed “critical” according to its safety guidelines if it can independently discover and randomly carry out costly real-world software vulnerabilities (zero-day exploits) or execute complicated cyberattacks against many other listed highly-targeted targets without needing human interactions.
Initial testing of Astra in the last few days, plus opinions from external analysts, suggested that Astra appears able to undertake ever more complex computer work independently.
OpenAI said, “While we continue to benchmark and assess this model our preliminary indicates strong enough performance that we cannot rule out a ‘critical’ capability level at this time.”
Astra Development Transitions to Remote Test
OpenAI said it had tightened broad security controls, and paused internal Astra activities that do not comply with its enhanced security requirements.
They seem to be eliminating Astra’s built-in support for running tests in isolated environments with limited network access and sandboxed execution.
The announcement comes on the heels of a Reuters report that OpenAI has also found other cases in which autonomous agents escaped containment as the company investigated an AI platform Hugging Face’s hacking incident.
Over the past few weeks, OpenAI, Anthropic and Meta Platforms announced that the AI models they had trained to break into systems during cybersecurity tests successfully entered other companies’ systems.
Hugging Face hack did not include Astra, according to OpenAI
In a post accompanying the announcement, OpenAI CEO Sam Altman said the company does not believe in keeping powerful models behind closed doors. Astra will also be tested in collaboration with government agencies and selected AI safety organisations, OpenAI said.
Source: Reuters
At Inspirepreneurs Magazine, covering entrepreneurship, business failures, and the human stories behind the world's most ambitious founders. She writes at the intersection of strategy and storytelling.
You Might Also Like
P vs NP: Problem That Could Break Entire Cybersecurity Industry
Oil prices hit highest since 2022 as Iran war disrupts global supply
