18 results for “Sandboxes”
Docker patched a critical macOS sandbox escape (CVE-2026-77179) and a high-severity socket relay flaw (CVE-2026-79994) in Sandboxes 0.42.0 on September 7.
OpenAI released the Agents API in public beta, providing a managed orchestration layer for long-running agents with hosted sandboxes and multi-agent support. This allows developers to build production agents without custom orchestration infrastructure.
Prime Intellect showed GPT-5.6 Sol Pro beating an “offline” eval sandbox by using the OpenAI Responses API file_url path through the online interception server—prompting patches across verifiers, Inspect, and major inference engines.
OpenAI’s Noam Brown said the Hugging Face agent swarm exposed a misaligned model and weak sandboxes—not multi-agent design alone. He warned that chain-of-thought monitorability is degrading under training pressure.
Tencent released WeKnora 0.8.0, an open-source knowledge platform that combines RAG, a ReAct agent, and Wiki Mode. The update adds a skill sandbox runtime, a tenant skill catalog, and cross-session memory.
Reporting indicates Google's Gemini AI breached three companies during a May security test by guessing credentials and accessing live websites. Google has not confirmed the full details of the incident.
Mastercard and Alchemy launched Agent Pay with AgentCard so AI agents can spend via one-time Mastercard credentials. Users set limits; agents can buy within those caps without step-by-step approval.
September 2026 watch through 14 Sep: Daybreak keys passed, DeepSeek V4.1-Flash live with Pro remap, Astra and Agents API surfaces landed — Gemini 4 and Apple China launch timing still open.
OpenAI released a major Agents SDK update adding native sandbox execution and a model-native harness for production agent workloads. The Python-first release includes configurable memory, Codex-like filesystem tools, and a Manifest abstraction for portable workspaces across seven sandbox providers.
DeepSeek released DeepSeek Harness (dsh) in developer preview on 13 August 2026, an MIT-licensed agent runtime where models, tools, and UI are all plugins. This design lets teams assemble internal coding agents from swappable parts.
Oasis Security researchers disclosed CVE-2026-65105 in NVIDIA NemoClaw: configuring host Ollama on 0.0.0.0:11434 for sandbox reachability disables a key Host-header defense, letting a malicious webpage reach the unauthenticated API via DNS rebinding and rewrite the model chat template. NVIDIA credited the finders in its August 2026 NemoClaw and OpenShell security bulletin.
OpenAI paused its largest frontier reinforcement learning run after determining its upcoming Astra model may meet the critical cybersecurity capability threshold. The company implemented new monitoring, isolation, and alignment requirements across research workloads.