CharlieKey takeaways from your favorite podcasts
← 31 October 2026The Cognitive Revolution

Mistral Large 4: The West’s Best Open Model? | AI Cyber Threats

A $6.3B Binance hack, AI agents attacking companies for cents, and Mistral's open trillion-parameter Large 4 frame an argument that cyber defense, open weights and compute efficiency now hinge on people and process.

Watch on YouTube ↗

Charlie's summary

  1. The Binance hack was old-fashioned phishing plus a human holding the keys, not a new AI-era attack

    Pash says an engineer's laptop was compromised, most likely by fake-interview malware, and attackers drained the hot wallet. They left enough Ethereum behind that normal transactions kept flowing, and pledged funds to a DeFi exchange so freezing Tether couldn't undo it. His point: “someone's got to hold the keys,” so crypto security always comes back to people. Pash also says the large amount was visible to everyone, unlike a stealthy superintelligence scenario.

    ▶ 2:55
  2. Agentic attacks in Southeast Asia cost 5 to 10 cents per compromised company

    Michael Freeman says that in the past two weeks his team saw 3 to 4,000 successful attacks by AI agent frameworks, which are being forked and extended quickly. They pivot, harvest credentials from ticketing systems, and attack Active Directory. They are still noisy, but they adapt and learn environments well. He had predicted in July that scripted-looking agent noise would turn sophisticated within months, and says it has.

    ▶ 33:59
  3. Offensive AI capability was trained in, and attackers' harnesses matter as much as the models

    Michael rejects the idea that LLMs simply stumbled into hacking. Models are trained on cyber ranges and CTF data, and attackers keep improving their harnesses day by day. Many run downloaded open-weight cyber models on local machines. His team built a small vulnerability-finding model for a couple hundred thousand dollars two years ago, so he calls reproducing this trivial. He thinks it can't be reeled back in and will be an arms race.

    ▶ 39:36
  4. Business processes, not software flaws, create the vulnerabilities that patching and AI won't eliminate

    Michael says AI researchers treat security as software bugs, but the worst holes come from lazy human processes. Examples: customer production API keys stored in support tickets, and a help desk giving everyone the same default reset password. Finding millions of vulnerabilities doesn't help when a Fortune 500 company needs years to test and patch. He thinks vulnerability zero will arrive, but slower than people claim. He also distrusts automated AI patching on OT systems until success rates reach 100%.

    ▶ 54:58

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.