Tensorlake

We built an AI security engineer with Tensorlake sandboxes that: - Scans any GitHub repo for SQLi, XSS, SSRF, and auth-bypass bugs - Runs 4 specialist Claude detectors in parallel, then triages false positives - Generates patches and validates each one against the project's real test suite - Runs every untrusted operation — clone, dep install, patch apply, test run — inside a Tensorlake sandbox, never on the host Built with @Tensorlake & @AnthropicAI link to code: github.com/tensorlakeai/e…

New guide to create massively parallel browser infrastructure with Chrome Devtool Protocol (CDP) and @tensorlake sandboxes. 1. Create a sandbox with the ubuntu-vnc image, which has Chrome pre-installed and also VNC if you want to record the browser sessions. 2. Start chrome in the sandbox with CDP enabled. 3. Create a tunnel with an API call. Tunnels are created over websocket so no additional software is needed to interact with Chrome. 4. Point your Agent to the local CDP port and make it drive the remote browser. 5. The CDP MCPs also work with this approach, so you don't really need to even write any code to make Claude or other agents interact with the browser. Alternatively, you can use Computer Use patterns which uses commands over VNC to operate the desktop. Tensorlake can run 1000s of sandboxes in parallel in a single project which enables creating Browser as an API services! link to docs below 🧵

Context are getting longer but Pi coding agent shows that you do not need to abuse them. In contrast to other coding agents, Pi keeps the system prompt under 1,000 tokens, which is 10x less than other harnesses like Claude code Learn more: tensorlake.ai/blog/pi-coding…

Is the real bottleneck for AI agents the model—or the harness? Terminal-Bench 2.0 suggests it might be the latter. ForgeCode ranks #1 among open-source harnesses, showing how much performance you can unlock without changing the model—just by improving how it uses tools. In ForgeCode’s case, the gains come from better tool orchestration and execution. Learn more: tensorlake.ai/blog/forgecode…

Computer use in @tensorlake sandboxes is live! We built isolated, reproducible desktops for agents — because the real problem isn’t the model, it’s the environment. github repo: github.com/tensorlakeai/c… tensorlake.ai/blog/building-…

Dynamic Resource Allocation of sandboxes have landed. You can now request for any CPU, Memory or Disk resources during sandbox creation requests. This happens in 300 milliseconds, not any longer than creating sandboxes with default limits. Dynamic resource allocations work with even custom sandbox images!

Tensorlake is an official environment in @harborml 0.5.0. Install the latest Harbor: pip install harbor[tensorlake] Run your first Harbor eval on Tensorlake sandboxes! If you’re using Harbor to publish research, we’d love to support you by offering credits on Tensorlake! tensorlake.ai/blog/tensorlak…

We just updated our docs. ✨ Check it out and get started with your first sandbox. docs.tensorlake.ai/sandboxes/intr…

Suspend = pause and resume the same sandbox Snapshot = save state and reuse it to create new sandboxes Many providers use these terms differently — which is where confusion starts. Tensorlake supports both, so you can choose the right behavior for each use case. tensorlake.ai/blog/suspend-v…

We're in. May 9th. SF. Come build with us — Tensorlake sandboxes included 😎 luma.com/rshibq6i

We're excited to accept the invite to the 2026 Scale Invitational! Packing sunscreen and 100k sandboxes😇

Inspired by @karpathy's Auto Research repo, we built an example for running auto-research loops on Tensorlake sandboxes! docs.tensorlake.ai/sandboxes/agen…

We shipped @tensorlake skills! Give Claude and Codex the ability to sandbox your agent harnesses, and dynamic abilities to wake them up periodically or when they receive a request to their APIs.

Scaling infrastructure for developer workflow is in the minds of many teams. I get asked about why would someone need sandboxes for running coding agents. Here are my thoughts on this: 1. Developers run at least 7-8 parallel coding sessions now. Managing git work-trees is a pain. If you are using Rust for ex, debug builds across all work-trees are going to blow up to 700-800GBs easily. Sandboxes solve this problem - spin up 100s of sandboxes and make them tackle issues in parallel. 2. There is no reason for a developer environment to be up when the coding agent finishes it work and waiting for more instructions. Sandboxes can be suspended and resumed without losing any in-memory or file system sate. 3. Coding agents do a lot of I/O with the model layer, and the harness pauses while the model is thinking. This means CPUs are not really getting used all the time. Sandboxes can be overprovisioned to get better resource utilization. We built @tensorlake sandboxes to live forever vs timing out sandboxes after a maximum 24h like most other sandbox infrastructure companies do. We went to lengths to implement "warm migration" so that we can move sandboxes to healthy machines without any disruption to agents running inside sandboxes.

Need 50 parallel environments from the same starting state? In tensorlake you just clone a running sandbox. We built it for RL workloads but teams end up using it for all kinds of things.

A core problem in RL evals is reproducible environment setup at scale. We made clone a core primitive in @tensorlake sandboxes. Copy the checkpoint in a sandbox, set up the agent harness, tools and clone and run N identical environments with different seeds. Cloning is very efficient, a content addressable file system only pulls blocks of the file system not present on a machine.

when a tensorlake sandbox goes idle, we snapshot it. next request comes in, it wakes up right where it left off. the agent has no idea it was ever paused, which is kind of the point.

We’re working on Sandboxes @tensorlake. We built a high-performance file system to help agents run tool calls faster for coding and data analytics workloads. The goal was near-SSD speeds inside sandboxes. To get there, we forked Firecracker and built a custom, block-based overlay filesystem with dirty bitmap tracking for fast snapshots. Here’s a benchmark measuring raw SQLite performance across sandbox providers. Tensorlake is 1.2–1.3x faster than Vercel, 1.5–1.9x faster than Modal, 1.6–1.7x faster than E2B, and 1.8–2.2x faster than Daytona.