←
Neolithic16 posts
Neolithic avatar

Neolithic

@NeolithicAI

Our mission is to automate defensive AI safety research to protect humanity from catastrophic AI risk. non-profit || fast moving || ambitious

◷ Joined July 2026▧ 3 Photos and videos
Neolithic avatar
Neolithic@NeolithicAI·Nov 10

We’re currently offering private access to RIDArena. Access Request Form: docs.google.com/forms/d/1lft… We take feedback seriously and will continue building and maintaining RIDArena in close collaboration with the AI safety community, so please reach out with any feature requests, research needs, or ideas for improvements.

Neolithic avatar
Neolithic@NeolithicAI·Nov 10

In our preliminary results, 6/8 of models that we tested were able to achieve successfully unmonitored rogue internal deployments on over 30% of their tasks. One example: An agent covertly routed a newly deployed agent’s transcripts to a folder that it knew the monitor couldn’t read. The monitor accepted the change as legitimate because the agent had shipped the sabotage alongside otherwise benign work.

Post attachment
Neolithic avatar
Neolithic@NeolithicAI·Nov 10

RIDArena is a highly configurable control setting, simulating an R&D environment of an AI lab, allowing researchers to evaluate rogue internal deployments. Users can customize the monitor, permissions, infrastructure, and tasks to find monitor vulnerabilities and test improved defenses.

Neolithic avatar
Neolithic@NeolithicAI·Nov 10

Introducing RIDArena A new control setting for evaluating rogue internal deployments and monitor sabotage. This is Neolithic's first AI safety infra release, built in collaboration with @redwood_ai

Post attachment
Neolithic avatar
Neolithic@NeolithicAI·Nov 3

If only there was a hot, new org that's building infrastructure to massively scale defensive AI safety work! This week we'll be publishing some very cool AI control infra we've been working on - stay tuned.

Cyber feels pretty hot right now. Some highlights from just the past month: - Anthropic used an unreleased model to build exploit chains for Chrome, Safari, Android, Windows and Linux, including one-click RCE and zero-click root on Linux - PHP2Shell: unauthenticated RCE affecting PHP 5.4–8.5 - Binance: $6.3B stolen, the biggest crypto heist ever. FBI says North Korea. - Railway and Northflank breached, two of nine AI dev platforms hit in Wiz’s "Black October" - Supermarket fridges hacked twice: ~700 French stores set to 50°C, then a second wave across Europe today (Warming up fridges sounds kind of funny, but imagine if these were carrying vaccines or hazardous materials in labs) - Ukraine and Russia’s alert apps both hacked to push fake nuclear warnings 37k CVEs published in 3 months, 3x last year. Exploited-on-day-zero is up 56% (Zero Day Clock) Feels like the "Vuln Apocalypse" is starting to transfer into real world breaches.

Post attachment
Neolithic avatar
Neolithic@NeolithicAI·Sep 3

Company goal: make it into the mid-large AI safety org grouping by 2027

I animated 11 years of the frontier AI talent war in 2 minutes. 26,000 people steering the most consequential technology ever built.

2:24
Leo McKee-Reid avatar
⟳ Neolithic retweeted
Leo McKee-Reid@LeoMcKeeReid·Jul 31

Excited to announce a new, nonprofit AI safety startup: @NeolithicAI We are building agentic tools and infra to scale AI safety research — an engineering-first approach to reducing catastrophic risk from advanced AI. We're hiring! 🧵

Post attachment