Meet Deriv One, our open-source enterprise AI platform for building agents, skills, automations, and internal apps—with built-in access control and sandboxing.
Bring your own models, run it on your infrastructure. ⚡️
Here’s how we built it. https://t.co/ZkM5xiSz0z
Your vendor risk platform costs USD 100,000 a year.
It works, but nobody on your team can tune it.
That inflexibility used to slow down our security team.
We dropped the vendor and built an in-house platform to score 400+ suppliers daily.
Here is how we built Vendiant:
https://t.co/Vo7oR4jXPw
A number goes up. Everyone celebrates.
Then someone asks "Why?", and nobody knows.
That question used to cost our analysts their whole morning.
Here’s how we got it down to minutes: https://t.co/0sohjLuZRn
BREAKING: Ram Altman has been spotted free-soloing the legendary 'Silicon Horn' to secure the last remaining cache of unallocated memory on Earth.
He stated in a blog post that once he reaches the top, AGI will finally have enough compute to process his latest funding proposals.
One process fails. The whole app goes down.
That was the risk of running everything on a single VM.
We moved to Kubernetes so a failing process wouldn’t take everything else with it.
Here’s how we made the move: https://t.co/wbUSaUDE0r
The worst legal AI failure isn't a crash.
It's a clean, confident answer that's wrong.
The usual fix? Tweak the prompt. Add a specific rule. Rerun a few examples. They die when the law, the country, or the model changes.
Here is how we do it: https://t.co/ohkaqGyg1F
You are describing abandoned software, not disposable software.
An unassigned blast radius only exists when code outlives its task. A script run once on a laptop to clean a CSV has no latent liability sitting in wait—the task executes, the output is verified by the operator, and the runtime dies.
The moment you ask "at what age does this output stop being usable by someone else?", you have already defined a durable system. You have an asset that persists across time, an asynchronous consumer, and a requirement for drift detection.
The birthday tracker’s number is the duration of the party. The payment system’s number is an SLA. The difference is that one was designed to die before an unassigned dependent could ever arrive, and the other was built to survive them.
AI makes code cheap. It doesn’t make failure cheap.
A birthday tracker and a payments system need very different levels of rigor.
Four questions help you decide whether your software is disposable or durable: https://t.co/VVmxFql0ya
You price both: Time Unnoticed × Blast Radius.
The birthday tracker and the payment system differ because the tracker's blast radius is zero. It can sit broken for years without harming anyone. In durable software, silent time compounds into catastrophic debt.
That’s why test suites and linters exist: to bring detection time down to 0 seconds on your local machine before the multiplier ever kicks in.
The execution record is its own store. Every exchange is saved as a row: who invoked the agent, the prompt, the reply, which tools ran, a hash of the code, and the names of the secrets that were loaded. Access grants keep their own history beside that. Memory changes are written as append-only files in the blob store.
Someone outside the platform team reads those records directly, or copies them from a read-only database role into storage they hold. What the agent posted to Slack is also visible in Slack itself.
Meet Deriv One, our open-source enterprise AI platform for building agents, skills, automations, and internal apps—with built-in access control and sandboxing.
Bring your own models, run it on your infrastructure. ⚡️
Here’s how we built it. https://t.co/ZkM5xiSz0z
@EvanKirstel@TechImpactTV Thanks @EvanKirstel, spot on about CISO concerns around agent security. We'd love to jump on the show and break down how we built the access control and sandboxing layers. Dropping you a DM with details!
Use the platform for small facts the next run should see. Put anything larger in your own storage.
Nothing written inside the container survives. Each run gets a fresh filesystem, and the container is destroyed when it finishes. The schedule row only keeps operational fields: last run time, status, error, and run count.
The built-in cross-run store is agent memory. A scheduled run loads the agent's existing memories into the prompt, and after the run the platform extracts short user-scoped facts and saves them for the next one. Each value is capped at about 2 KB. That is the place for "remember this." It is not a file store or a cursor database.
Documents you want searchable later go in the agent's knowledge base, which lives in the platform blob store. Structured state — cursors, datasets, last-processed ids — belongs in your own storage. The sandbox has no network unless that run is explicitly granted it, so reaching that store is a deliberate permission, not something the container has by default.
Both, but the boundary that matters by default is filesystem and network isolation.
Model-written code runs in a throwaway container: network off, read-only root, all capabilities dropped, no credentials mounted, and the container is destroyed after the run. Set SANDBOX_RUNTIME=runsc if the host has gVisor. Scoped API permissions do not help if that code can still open a socket or read the host.
Agent identity is a separate boundary. Each invocation is its own process and only that agent's secrets are loaded into it, so a prompt injection against one agent cannot read another's tokens. Who is allowed to invoke an agent is an access grant, not the sandbox.
Other Statistics:
Claude Opus 5.5: 9.3 mins, 2 turns
Muse Wonder 1.0: 1.5 mins, 2 turns
GPT-6.1 Sol: 4.8 mins, 15 turns
Gemini 4 Argon: 3.6 mins, 2 turns
Want to run the experiment yourself, here is the prompt we used:
"""
Build a single self-contained `index.html` in the current directory: a Flappy Bird-style game where a little Deriv token flaps through the gaps of a live candlestick trading chart, made with three.js.
Make it feel like a polished, playful game skin on top of a real Deriv trading terminal: dark chart background, crisp grid, the Deriv red brand colour (#FF444F) on the player and logo, green (#4BB4B3) and red (#EC3F3F) candles.
Requirements:
- One file only. Load three.js (and addons, if you need them) from `https://t.co/JrFPqXshqF` via an import map. No other external assets: no textures, models, sounds or fonts to download. Generate everything procedurally (system fonts only).
- Gameplay: classic flappy loop. Gravity pulls the player down, each flap gives an upward kick. The "pipes" are candlesticks: each obstacle is a pair of tall candles, one hanging from the top and one rising from the bottom, with wicks, and the player must fly through the gap between them. The gap follows a believable random-walk price path, so the course reads like a moving market. Passing a pair scores +1. The scroll speed rises slowly and the gaps tighten a little as the score climbs. Touching a candle body or wick, or the top/bottom of the chart, ends the run.
- Player: a glossy round Deriv token in #FF444F with a white lowercase "d" on its face. It tilts up when it flaps and noses down as it falls, squashes and stretches on each flap, and leaves a short glowing trail like a price line behind it.
- World: a dark trading chart that scrolls right to left, with a subtle grid, a price axis on the right whose labels tick as the chart moves, and a thin live price line tracking the player's height. Candles are crisp, slightly beveled 3D boxes with wicks. Their colour (green or red) comes from whether the price went up or down from the previous candle. A small **"deriv"** wordmark in the top-left corner in the brand red, plus a ticker such as "Volatility 100 Index" with a live-looking price. Tiny details count: soft bloom on the price line, a faint depth parallax of background candles, and a "+1" pop when a gap is cleared.
- Controls: Space, Arrow Up or W flaps (click or tap also flaps). Arrow keys and WASD both work. Show the controls in a small hint on screen.
- Starts instantly: gameplay is running on page load, or starts on the first click or key press. No menus, no "press start" screen that blocks play.
- Attract mode: if there is no input for 1.5 seconds, a simple built-in autopilot plays the game convincingly by flapping to stay centred in the next gap. Any input hands control back to the player immediately.
- Never stalls: on game over, show the score briefly (styled like a trade result, e.g. "Position closed: 7") and restart automatically within 1.5 seconds.
- A score/HUD that fits the style, plus juice: particles on each flap, a red flash and small screen shake on a crash, and a smooth camera.
- No pointer lock, no fullscreen requirement, no sound before the first user input.
- It should run smoothly at 60 fps on a laptop and must have no console errors.
Write the file and stop. Do not ask questions.
"""
Muse wonder dropped the ball wonderfully when we compared with Opus 5.5 & Gemini 4 Argon. Same prompt.
@finkd Talk to your people, they need to step it up 🙃
6.1 Sol wasn’t even in the race, it got lost on the drive to the track.
Claude Opus 5.5: $1.5
Muse Wonder 1.0: $0.36
GPT-6.1 Sol: $0.18
Gemini 4 Argon: $0.24