Skip navigation
AI security stack is getting f...cking important
0:00 / 0:32

AI security stack is getting f...cking important

Promix

@Promixyz1.3K followers

Visit on X
543 likesDownloadOriginal post

27,800 views

AI security stack is getting f...cking important 10 open-source GitHub projects for letting agents use real tools without giving them the keys to EVERYTHING 01 E2B ▸ → isolated sandboxes for agent execution 02 gVisor ▸ → sandbox untrusted workloads ISOLATE THE AGENT 03 Firecracker ▸ → lightweight microVM isolation 04 Kata Containers ▸ → containers with VM-level isolation CONTROL THE TOOLS 05 Open Policy Agent ▸ → policy checks before actions execute 06 Casbin ▸ → permissions + access control PROTECT THE SECRETS 07 Infisical ▸ → secrets management for AI infrastructure 08 SOPS ▸ → encrypted secrets inside repos WATCH WHAT HAPPENS 09 Falco ▸ → detect suspicious runtime behavior 10 Trivy ▸ → scan code + containers + infrastructure the security loop: request → permission → sandbox → tool → execute → inspect → allow / block → log 3 layers I'd actually build: agent: E2B → OPA → Casbin runtime: Firecracker → Falco → Trivy secrets: Infisical → SOPS and this is where it gets interesting: giving an AI agent more tools makes it more useful it also gives it more ways to break something the next agent stack isn't just memory + tools + reasoning it needs permissions ⭣

27 Comments

Comments from the original post on X (Twitter)

  • Karpacho's profile picture

    @Karpachooa

    GitHub allows to do more than enytime Do you agree @grok?

    3 likes

  • Lummox's profile picture

    @Lummox_eth

    Another group of alpha repos

    2 likes

  • Dimitrios Prasakis's profile picture

    @dimiprasakis

    You forgot the easiest to use sandbox in the market for Claude Code and Codex:

    2 likes

  • maximus's profile picture

    @maximus__m

    Added

    1 like

  • Nguyễn Chí Trọng Nghĩa's profile picture

    @hoatanhoano

    Đúng cái mình đang tìm, kết quả đẹp trên bảng đo chưa chắc đúng với việc thật của mình . Có cách nào rẻ hơn không bác?

    1 like

  • Colbert's profile picture

    @sirshibaninja

    Security is going to be the biggest bottleneck for agent adoption. Love seeing these tools getting highlighted, keeping things sandboxed is non-negotiable at this point.

    1 like

  • Stanislav Sorokin's profile picture

    @stas_sorokin_

    The safest key is one the agent never holds: 27,205 of our decisions last week ran on a local model, never touched an API, cost 0 dollars. Where it hands off is the real design question. Which of the 10 do you run?

    1 like

  • Akash Hadagali's profile picture

    @akashc777

    Sandboxes cap the blast radius, but they don't decide whether this call, for this user, right now, should run. That needs a live scope check per tool call and an audit row written before execution.

    1 like

  • Sharapov Andrei's profile picture

    @AdrasSad

    On the agent layer, do you run OPA before every tool call, or only before irreversible ones? A policy check on each step looks right on paper, but in a multi-step loop it can stall the agent harder than the sandbox itself.

    1 like

  • AI Desk Club's profile picture

    @aideskclub_

    The missing step is the deny path: when policy blocks a tool call, route it to a human approval queue instead of letting the agent retry until the sandbox becomes a suggestion.

    1 like

  • Payment Sandbox's profile picture

    @BritJonovski

    sandbox + scoped creds is the correct order. e2b for isolation is fine, but if the agent can still pass a wildcard API key through, the sandbox is theater. scoping first, then exec.

    1 like

  • Neeraj Kumar Singh Beshane's profile picture

    @i_am_neerazz

    The gap in that loop is credential scope. If the sandbox gets a long-lived API key, OPA can approve one call but the key still reaches its full scope. Test: from inside the sandbox, call an API the task never needed. If it works, you isolated the process, not the authority.

    1 like

  • Kshitiz Loharuka's profile picture

    @KshitizLoharuka

    Hey check out Have been using for a while now

    1 like

  • s0sta's profile picture

    @s0sta_

    A security stack list is more useful when ordered by what to fix first. For most teams that order is identity, then secrets, then model input handling, then everything else.

  • _alphashark_'s profile picture

    @_alphashark_

    sandboxes don't stop the biggest agent security failure: tools still inherit broad oauth scopes. capability tokens and per-action policy checks matter more than process isolation.

  • spenz's profile picture

    @SpenzDigital

    the gvisor one gets all the credit but ephemeral creds do the quiet work. sandbox stops the escape, scoped keys stop the exfil.

  • Priyansh Dugar's profile picture

    @_Priyansh_10

    great list. one that covers the control the tools + protect the secrets + watch layers in one place built for agents specifically: intercepts MCP/langchain/coding agent tool calls allow/block, policy, redacts secrets before the model, audits logs.opensource

  • Open Source Daily's profile picture

    @OpenS_Daily

    E2B sandboxes plus gVisor can give defense-in-depth for agent tool use.

  • The Tx Queue's profile picture

    @feeratefield

    isolated execution + no ambient creds. only combo i trust for anything touching a real tool.

  • Rose's profile picture

    @frogbetrose

    Sandboxing agents is smart. Like must-hits, you cap the downside. What's the rake on that stack?