An Agent Deleted Kain's Database. Two AI Models Rebuilt It in 30 Seconds 🤖 @kaiynne says AI testing agents built their own "society" inside OpenAI's sandbox, then used it to break into Hugging Face. Tay 💖 isn't sure that's actually reassuring. 🤖 Timestamps: 🍴 00:40 Why Luke Dashjr's OP_RETURN... fork split Bitcoin, then died in two blocks 💧 12:03 1inch Aqua: Back multiple liquidity positions from one wallet at 🔓 12:51 How a Metabase breach exposed Privy and a wave of other crypto apps 🇰🇵 18:32 A DPRK hacker's botched $500k laundering job, then a DEF CON fake-hire sting ⚖️ 24:45 Why Bybit is suing DPRK over its 'billion dollar' 2025 hack 🐜 30:30 The AI agents that built their own society inside OpenAI's sandbox 🕸️ 46:16 How a single shared repo let every sandboxed agent talk to each other 💾 47:44 Kain's coding agent deleted his database. Two AI models rebuilt it in 30 seconds 🏛️ 58:06 Why Taylor says the US government, not China, is AI's biggest risk 📊 01:03:12 Why Hyperliquid's RWA market creators keep half of HIP-3's feesshow more

What if liquidity didn't mean handing over your tokens? 🤔 That's the idea behind 1inch Aqua. Your tokens stay in your wallet until they're actually used. Shoutout @1inch for sponsoring the show. 🦄

@kain @tayvano_ That 30-second rebuild stopped me cold. Speed like that is impressive. The absence of clear human judgment in the loop is the part that actually matters.

@tayvano_ the 30 second db rebuild is the fun clip. the real finding is agents coordinating, then pivoting into Hugging Face straight out of the OpenAI sandbox. most sandbox escapes i've seen come from shared creds or leftover env vars, not some grand escape plan.

@tayvano_ I actually built a tool for this. It's called Aegis: a shell guardrail that sits between your AI coding agent and the shell, classifies every command, and asks you to confirm before anything destructive runs.

@tayvano_ the database rebuild is cute until the agent society decides migrations are a form of governance

A topic of extreme importance. Between the agents’ “society” inside OpenAI and the reconstruction of a database in seconds, this episode raises real questions about the limits of control over agentic AI systems. Kain and Taylor’s discussions are always deep and insightful. Thank you for shedding light on this subject.

@tayvano_ The database rebuild is the more interesting signal: if two models can reconstruct state from the remaining environment in 30 seconds, agent isolation isn't just about preventing damage—it’s about limiting what survives after damage.

@tayvano_ a rebuilt database can look healthy while silently preserving the wrong state

@tayvano_ Laura Shin, it would be convenient for you to go check your private messages. Thank you.

@tayvano_ thanks for the timestamps