2
Retroactive funding: agent-glovebox, open-source sandboxing for AI agents
1
ProposalGrant
Closes December 13th, 2026$0raised
$500minimum funding
$65,836funding goal
Offer to donate
30 daysleft to contribute
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$65,836 left to reach the funding goal.
Summary
agent-glovebox is an open-source, security-focused sandbox for running evals on AI coding agents. It constrains what an agent can touch, reducing the damage from mistakes, misbehavior, and prompt injection. It also enables tamper-proof logging and careful customizable monitoring integration. I began working on it independently from May 24th to August 9th without funding. This project is retroactive compensation for that work.
About me
I’m Alex Turner, an AI safety researcher, formerly at Google DeepMind and a UC Berkeley postdoc. On August 10th, I began work at FAR.AI as a Visiting Engineer (which I wrapped up on November 6th).
What’s been built
agent-glovebox, a sandboxing layer for agent workflows, isolating agents from credentials, the network, and the host filesystem except where explicitly allowed. Tested by regular CTF red-team exercises using open models, Mythos 5.2.
agent-sanitizer for reducing prompt-injection attacks.
ci-truth-serum checks for common footguns in GitHub CI.
I developed agent-review and agent-resolve-merge-conflicts to allow high-throughput agent-driven development.
Why it matters
Rogue AI incidents are important to track, log, and monitor. Glovebox lets any AI alignment organization use a serious sandbox with minimal workflow disruption and engineering investment.
Status
In final stages of integration with METR's Hawk eval platform and FAR.AI's flamingo cluster, after which I plan to expand integration to more alignment organizations.
Funding
This is a retroactive award for completed work. Amount: $65,836. ~2.5 months * $25.5k / month = $63,250, plus $2,586 of expenses for experiments and coding agent subscriptions. I completed this work without waiting for funding confirmation, which seems correct given the Hugging Face incidents revealed two months later in July.
Conflicts of interest
None that I’m aware of.