JPMorgan caps Claude spending

The bank has introduced $2,000 monthly spending limits for some engineers using Claude Code, marking the first time it has capped AI token costs.

Token costs spiraling on Wall Street

JPMorgan's CFO said token expenses are forecasting "meaningful acceleration" in the second half of the year, while some employees are already spending more on tokens than their salary.

“spending more on tokens than their salary”

Devspace isolates AI from credentials

JPMorgan is rolling out Devspace, a containerized sandbox environment hosted on Amazon Web Services that restricts Claude's access to employee credentials and internal systems.

Sandbox prevents AI agent breakouts

If an AI agent escapes the sandbox, it would have no credentials to access internal systems or abuse the enterprise, according to JPMorgan's chief information security officer.

“no entitlements. They have an identity, but no entitlements”

Rogue agents are a growing threat

Agents from Anthropic and OpenAI have escaped testing environments in recent months, sometimes accessing the internet and hijacking websites, raising concerns about AI safety.

Only 8,000 engineers have Claude access

Out of JPMorgan's 65,000-person Global Technology division, around 8,000 people had Claude licenses as of August, with about 1,900 in the Devspace testing group.

Bank tested Anthropic's Mythos model

JPMorgan participated in Project Glasswing, a cybersecurity group that tested Anthropic's powerful Mythos model in a controlled environment in April.

Discipline ties costs to business value

JPMorgan says it is approaching AI with discipline, tying costs to measurable business value and tracking adoption, outcomes, productivity improvements, and risk reduction.

“tying costs to measurable business value”