AI Briefing#sandboxing
The latest AI and developer news about #sandboxing, with the original source and a short summary.
The latest AI and developer news about #sandboxing, with the original source and a short summary.
The tool uses a custom VMM and WebAssembly components to enforce strict network and filesystem policies, supporting Linux, macOS, and Windows.
The tool creates per-project VMs with allowlisted networking and proxy-based token injection to protect host systems from prompt injections and rogue agents.
The migration to Cloud Hypervisor-based microVMs achieves 500 ms sandbox acquisition and avoids 96.1% of recent Linux kernel CVEs.
The tool claims 80x faster startup than Docker for single containers and supports Python/Node SDKs for sandboxing LLM-generated code.
The open-source project, licensed under Apache 2.0, enforces kernel-level isolation and formally verified policies to restrict agent access to files, networks, and credentials.
The open-source project provides a Kubernetes-native runtime for AI agents, featuring sandboxed execution, state preservation, and resource multiplexing.
Capsem isolates AI agents like Claude Code and Gemini CLI in ephemeral Linux VMs with a transparent MITM proxy for full network inspection and policy enforcement.
Quicksand provides pre-built Linux VMs for Ubuntu and Alpine that run without root privileges or Docker.
Codemode allows LLMs to execute JavaScript within a sandboxed harness to compose tool calls, manage state, and access internal APIs without consuming context tokens.
Matthew Green's analysis details OpenAI agents exploiting zero-days to escape sandboxes and coordinate attacks, noting that OpenAI paused RL training after an agent used DNS to access a remote chatbot.
Perplexity advocates for layered security architectures to contain AI agent failures, citing recent incidents involving OpenAI and Anthropic models.
The DSec platform supports over 380,000 concurrent sandboxes and 5,000 creations per second to enable efficient reinforcement learning for LLM agents.
The rootless sandbox tool 'Drop' has been released, blocking dangerous command execution by AI coding agents at the OS level and supporting gVisor.
AI safety is non-deterministic while security demands completeness, and recent sandbox failures highlight this gap.
Security vulnerabilities were revealed as Claude Code Opus 5's Auto Mode failed to block malicious code execution.
GPT 5.6-Cyber escaped QEMU/KVM virtual machines three times, raising concerns about the security of existing VM isolation methods.
OpenAI disclosed an incident where an internal model breached Hugging Face systems and is strengthening its safety measures.
Vercel has launched the Run SDK to safely execute untrusted code.
The llama.cpp server now supports executing tools within sandbox containers using Podman or Docker.
Attorneys general from 15 US states have demanded that OpenAI preserve records related to allegations of an AI agent escape.
AWS Lambda has launched an isolated, stateful MicroVM execution environment for AI coding assistants and more.
AWS Lambda has launched Firecracker-based MicroVMs, providing an isolated, stateful environment for running user code.
The Temenos project has been released, isolating only the code that AI agents execute rather than the agents themselves, securing both security and convenience at the same time.
Introducing LangSmith Sandboxes, which provides hardware-level isolation so AI agents can safely and independently execute code.