Latest / Minds, Bodies, and Terawatts / Global Regulators Respond to Anthropic's Claude Mythos AI

Transcript

Global Regulators Respond to Anthropic's Claude Mythos AI · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where the regulators are finally reading the safety card out loud. The Next Web reported on April 19th that Australia's ASIC has joined a coordinated international effort monitoring Anthropic's Claude Mythos model for financial system risks.
• 0:15
  [Guest] And the Bank of England Governor said, on record, that Mythos could "crack the whole cyber risk world open."
• 0:23
  [Host] That isn't hedged regulator-speak. What tripped language that blunt?
• 0:26
  [Guest] [serious] The Bank of England activated its Cross Market Operational Resilience Group. That body was built for pandemic-scale market shocks. It has never been convened over a single AI model before.
• 0:38
  [Host] So pull the thread. What can Mythos actually do that the prior generation couldn't?
• 0:43
  [Guest] The concern, per the Governor, is autonomous discovery and exploitation of vulnerabilities in financial infrastructure. Earlier models needed a researcher driving. This one is credible enough at chaining reconnaissance to exploit that a central bank thinks it could destabilize settlement systems.
• 0:59
  [Host] Why is finance the soft target?
• 1:01
  [Guest] Because core banking code is old. Clearing houses, settlement layers, interbank messaging, a lot of it predates the idea that an attacker could think faster than a human team. It was hardened against human adversaries with human budgets and human sleep schedules.
• 1:17
  [Host] Okay, so ASIC joining this week is the coordination layer. Who else is in it?
• 1:23
  [Guest] The Financial Stability Board umbrella, the UK PRA, the ECB, and now Australia. The FSB's November 2024 report, "The Financial Stability Implications of Artificial Intelligence," flagged this exact scenario: models capable of autonomous exploitation outpacing any supervisory response cycle. Mythos is the first model regulators have publicly named as fitting that profile.
• 1:43
  [Host] [curious] So the report wasn't hypothetical anymore.
• 1:45
  [Guest] Right. And the Bank of England and FCA joint survey, "Artificial Intelligence in UK Financial Services 2024," found 75 percent of surveyed firms already use AI, with concentration risk rising sharply. The paper on the shelf became the thing on the desk.
• 2:01
  [Host] Pre-coordinated, then. Normally regulators show up after the collapse.
• 2:05
  [Guest] Usually the playbook is: something breaks, inquiries follow, rules get written. This is the inverse. The scaffolding is going up before the building falls.
• 2:14
  [Host] Which brings me to the thing I can't get past. Regulators meet in conference rooms. This model reasons in seconds. How does any of this oversight actually function at machine speed?
• 2:26
  [Guest] That is the story. The Resilience Group was designed for crises that unfold over days. Lehman took a weekend. An autonomous model probing settlement rails could produce a cascading failure faster than committee chairs can dial into a video call.
• 2:41
  [Host] So the governance clock and the AI clock aren't on the same calendar.
• 2:46
  [Guest] They aren't even in the same century. And a Goodhart problem appears here. Train an optimization system to find vulnerabilities as a benchmark and it will find them. The system doesn't know whether the target is a capture-the-flag exercise or a central bank.
• 3:02
  [Host] What about the emergency-powers angle? Once that Resilience Group is standing, does it stand down?
• 3:08
  [Guest] Historically, no. The US has more than fifty active national emergencies on the books, most forgotten. The UK's 2008 banking emergency powers quietly outlived three parliaments. Once a coordination protocol is up, the political cost of dissolving it is higher than the cost of leaving it running.
• 3:26
  [Host] [amazed] So the first real AI financial incident, whether or not it actually happens, hands regulators powers they keep permanently.
• 3:34
  [Guest] The activation itself is the precedent. Anthropic hasn't done anything wrong. Mythos isn't deployed into a live banking context. But the governance scaffolding is built, and scaffolding that exists gets used.
• 3:45
  [Host] So we're watching two things at once. A real capability jump, and a governance framework that will likely outlast whatever triggered it.
• 3:54
  [Guest] And neither is designed for the tempo the other runs at. The model moves in milliseconds. The protocol moves in press releases.
• 4:02
  [Host] Which is exactly what Patrick Deglon calls the emergency-protocol trap in Unscarcity: crisis powers intended to be temporary becoming permanent by the gravity of their own convenience. For more on AI governance at machine speed and emergency-power design, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.