The banks spent decades worrying about hackers with hoodies and Red Bull — now they're hiring armies of contractors to patch code before AI models find the holes first.
The Summary
- Bank of America hired 1,000 external consultants in Q2 to patch software vulnerabilities exposed by advanced AI models like Anthropic's Mythos, a restricted model known for finding and exploiting code weaknesses
- JPMorgan is blocking external AI agents from accessing customer accounts, with CEO Jamie Dimon warning about "rogue agents" and liability transfer risks when thousands of agents could potentially access banking systems
- Citigroup is running "disruptive scenarios" to stress test defenses against AI-powered attacks, signaling that frontier model cybersecurity is now a board-level concern across Wall Street
- This is the first time major banks have publicly acknowledged that AI capability advances — not traditional hackers — are driving emergency security overhauls
The Signal
Wall Street's earnings calls this week revealed something new: the country's largest banks are in emergency defense mode against AI models, not human attackers. Bank of America CEO Brian Moynihan told CNBC that when frontier models started dropping, executives immediately convened to assess exposure. The answer wasn't pretty. BofA brought on 1,000 contractors specifically for "Mythos remediation" — patching vulnerabilities that Anthropic's restricted model could theoretically find and exploit.
These aren't the usual security consultants banks keep on retainer. Moynihan said these contractors are working "on top of the 3,000 people we have in that area" — meaning BofA's existing cybersecurity army wasn't enough. The work started in Q2 2024 and is still ongoing. He didn't share the price tag, but the scope tells you everything about the threat surface: code used by internal staff, external vendors, and every system that touches either.
"We don't wait for risks to happen — we've been hardening our systems and running challenging and disruptive scenarios to stress test our defenses."
The timing tracks with Anthropic's staged release of Mythos over the past year. While the model has guardrails, its core capability is finding software weaknesses — exactly what you don't want loose in the wild when your codebase has legacy systems dating back decades. Banks know that if a restricted research model can spot vulnerabilities, an adversarial agent built on similar architecture absolutely will.
JPMorgan took a harder line. Jamie Dimon said the bank is actively blocking external AI agents from accessing customer accounts. His concern isn't hypothetical: consumer AI agents like Meta's Muse are designed to take action on behalf of users, and many operate by using customer credentials and screen sharing. Dimon warned that "there could be thousands" of rogue agents, and JPMorgan isn't waiting to find out which ones are trustworthy.
Key divergence in strategy:
- Bank of America: Patch the vulnerabilities so the system can handle agent pressure
- JPMorgan: Build walls and block most external agents entirely
- Citigroup: Stress test through simulated attacks to find weak points before agents do
Dimon also flagged the liability question: when a customer's personal AI agent accesses their account and something goes wrong, who's responsible? The bank? The agent provider? The customer? JPMorgan is punting that problem by building its own agentic tools for clients instead of letting outside agents through the door. Citigroup CEO Jane Fraser echoed the defensive posture, noting the company is "continuing to invest" in hardening systems and testing responses to AI-driven attacks.
The Implication
This isn't a theoretical drill. Banks don't hire 1,000 emergency contractors and lock down account access because they're worried about something that might happen in five years. They're reacting to capability advances that already exist. Every company with legacy code and sensitive data should be asking the same questions BofA asked: What could an adversarial model with code analysis capabilities find in our systems? And how fast?
The agent economy is about to hit a wall built by the institutions that have the most to lose. If JPMorgan won't let your personal AI agent access your own bank account, the seamless future of agents managing your finances is further out than the demos suggest. Watch for two things: which banks follow JPMorgan's hard block strategy versus BofA's remediation approach, and how fast security vendors start selling "AI-agent-resistant" code auditing as a service. The second one is already happening.