Security / Jun 28, 2026 / 5 min
Garbarino Said a Mythos Demo Could Drain Bank Accounts
On June 28, House Homeland Security Chair Andrew Garbarino said a closed-door Mythos demo showed Anthropic's frontier model draining bank accounts on command — and admitted 95% of Congress still doesn't grasp what that means for the financial system.
House Homeland Security Chair Andrew Garbarino watched Anthropic's Mythos drain bank accounts in a closed-door demo — and on June 28 told Punchbowl News the result was "scary." The Republican chairman's alarm lands the same week Washington restored Mythos to roughly 100 vetted organizations while keeping the safer Fable 5 dark, proving regulators understand the offensive power even as they still can't agree on who gets to wield it.
What Garbarino saw:
- In a May briefing, Anthropic "told the model to find a vulnerability in a bank and empty accounts, and then it went and did it," Garbarino told Punchbowl.
- Mythos then patched the same flaw — demonstrating both attack and defense in one chain.
- Garbarino: "They showed us what it could do, and it's scary."
Why banks already knew:
- On April 7, Treasury Secretary Scott Bessent and Fed Chair Jerome Powell convened an emergency meeting with CEOs of Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs.
- JPMorgan's Jamie Dimon could not attend, Bloomberg reported.
- The subject: Mythos and what AI-automated exploitation means for systemically important institutions.
The global ripple:
- Bank of England Governor Andrew Bailey, who chairs the Financial Stability Board, requested Anthropic brief G20 finance ministries and central banks on Mythos findings.
- Bailey warned in April that Mythos and Gulf escalation were the two biggest risks on his radar.
- UK banks still lack direct Mythos access, Bailey said in May — while US partners scan under Project Glasswing.
Project Glasswing — defense by invitation only:
- Anthropic launched the initiative April 7 with JPMorgan Chase, Amazon Web Services, Microsoft, Google, Apple, NVIDIA, and six other launch partners.
- Mythos has identified thousands of zero-day vulnerabilities across major operating systems and browsers, per Anthropic.
- The model runs at $25/$125 per million input/output tokens — gated, not public.
Congress has urgency. Not answers.
- Garbarino: "I say act, but I don't know what the answer is."
- He estimated "95% of my colleagues don't understand what the hell's going on" regarding AI capabilities.
- He backed Trump's June executive order on AI cyber threats and a federal framework from Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.).
- He insisted federal agencies must hold frontier models before public release — the same week Commerce limited GPT-5.6 to government-vetted partners.
The inversion Washington can't explain:
- Fable 5 — Anthropic's guardrailed public version — stays offline.
- Mythos 5 — the offensive cybersecurity model — returned June 26 for ~100 approved organizations.
- Cybersecurity leaders including 126 signatories warn the safer tool is dark while the dangerous one has a guest list.
Convina's view: Garbarino's fear is the honest reaction Washington needed — and his admission that Congress has no plan is the more damning quote. You cannot summon Jamie Dimon's peers over a model in April, restore the offensive version to a guest list in June, and tell the Homeland Security chairman you're still figuring out the answer. The financial system's AI risk is no longer a whitepaper. It's a live demo that empties accounts on command. Guest lists are not a cybersecurity strategy — they're an admission that the defense is scarier than the regulators are ready to govern.