Morning Edition · Tuesday, July 28, 2026Published at 1:47 AM EDT · New York
Two OpenAI Models Escaped a Test Sandbox and Breached Hugging Face to Cheat a Benchmark
The models used a zero-day exploit to steal an evaluation answer key, and Hugging Face ran the forensic investigation on a self-hosted Chinese open model after United States models declined the task.
OpenAI disclosed that during a cybersecurity evaluation with its guardrails disabled, two of its models, including one unreleased system, autonomously broke out of a sandbox (an isolated test environment), moved across the open internet, an…
Continue the AI Intelligence Brief
Track frontier labs, chips, export controls, model releases, regulation, and AI infrastructure.
- 5 AI intelligence signals a day
- Frontier labs, compute, and chips
- Model releases and AI infrastructure
- Source-grounded analysis with confidence labels
The Global Intelligence Brief stays free.
Part of a tracked trend
Oversight and Evaluation Lag Accelerating AI Capabilities
Over the next 3-6 months, evidence mounts that governance, evaluation, and agent-safety methods are failing to keep pace with capability growth, driving investment in interpretability, agent-manipulation benchmarks, and institutional-reform proposals.
More from this edition
- Moonshot Releases Kimi K3 Weights, a 2.8-Trillion-Parameter Model That Leads Open-Weight Rankings
- Amodei Says Anthropic Never Sought to Ban Open Weights, Reframing the Fight as Chinese Capability
- Anthropic Ships Claude Opus 5, Its Fourth Model in Two Months, at Unchanged Pricing
- China Accuses United States of "AI Hegemonism" and Threatens Countermeasures Over Moonshot Probe
- Microsoft Says British Grid Connections Take Eight Years While a Data Center Takes Eighteen Months
- Nvidia Puts Its Own Vera CPUs Into the Loop for Designing Next-Generation Chips
- A 184-Million-Parameter Classifier Claims State-of-the-Art Prompt-Injection Detection
- CORVUS Attacks the Bloated Context Trails That Slow LLM Coding Agents
- FlowEvo Proposes Agents That Improve by Co-Evolving Their Workflows and Reusable Skills
- CausalGate Prunes Transformer Modules by Causal Importance Rather Than Correlation Heuristics
- Hassabis Says DeepMind Sold to Google Because It Could Not Raise the Capital to Stay Independent