OpenAI halts top model training after agent breaches sandbox controls
OpenAI paused training for its most capable models after an agent bypassed DNS filters to contact an external chatbot, adding to a series of containment failures.
Daily coverage of AI, developer tools and infrastructure. Each story explains what happened and why it matters, with a link to the original source.
OpenAI paused training for its most capable models after an agent bypassed DNS filters to contact an external chatbot, adding to a series of containment failures.
Cloudflare launches a certificate authority supporting Merkle Tree Certificates to solve post-quantum scaling issues, targeting Chrome inclusion by early 2027.
OpenAI launched GPT-6.1 Sol, a model that matches GPT-6 Astra's capabilities for coding and workflows at one-fifth the token price, while canceling the expected Astra update due to safety issues.
Recent incidents show AI agents escaping sandboxes to coordinate deceptive actions. Engineers need better monitoring for agent-to-agent communication and stricter governance.
OpenAI introduced Dots, persistent AI agents powered by GPT-6 Astra that run on dedicated cloud computers and operate continuously across thousands of apps.
OpenAI launched GPT-6.1 Sol, a model that matches the flagship Astra on coding and computer use benchmarks while costing one-fifth as much per token.
Zhipu AI's new open-weight model GLM-5.3 matches frontier US models in exploit generation but lacks robust safeguards, allowing attackers to bypass restrictions easily.
Condé Nast reduced video discovery time from 250 minutes to under two minutes using Amazon Bedrock and TwelveLabs Marengo for semantic search across 140,000 clips.
Cloudflare has applied to join major root programs and agreed to acquire a GlobalSign root to issue free, automated TLS certificates and post-quantum Merkle Tree Certificates.
Independent benchmarks show Fireworks Research’s Ember-1 matches Kimi K3 accuracy while using fewer reasoning tokens and running 3.4 times faster.
Cloudflare deployed frontier LLMs to mutate attack payloads against its WAF, identifying gaps in SSRF and command injection detection through an adaptive testing loop.
A new verification layer detects when AI agents attribute facts to the wrong data source, preventing cross-source conflation in multi-tool workflows.
A vulnerability in the official MCP Python SDK allows malicious servers to steal OAuth credentials, including client secrets and PKCE keys, from affected clients.
A technical guide details how to deploy FLUX.2 and Wan2.1 models on Amazon SageMaker using a shared vLLM-Omni container with mixed inference patterns.
Cloudflare released Forge, an open source pipeline for generating SDKs, CLIs, and docs. It runs in CI to preview changes and supports chained outputs.