How does JPMorgan Chase govern responsible AI?
JPMorgan Chase governs AI by matching controls to what a system can do. Its March 2026 engineering post says confined, read-only agents merit lighter guardrails while agents combining untrusted inputs, sensitive data and authority to act need continuous enforcement; global CIO Lori Beer said in October 2026 that the bank has put guardrails around agents 'at a very fine grain level'. J.P. Morgan Payments' data chief Zack Anderson wrote in September 2026 that payment agents need enforceable controls, not prompts: an independent policy engine that checks the agent, whom it represents and its mandate before execution. The Fence framework tests LLM use cases for hallucination, topic drift and prompt injection with synthetic data, and AI Research runs a firmwide Trustworthy AI Center of Excellence on explainability, fairness, robustness, security and privacy. The 2025 10-K, filed February 13, 2026, names AI failures, biased output, agentic systems acting on sensitive data, AI-enabled cyber threats and workforce displacement as risks. Externally, the bank blocks outside AI agents that use customer credentials, which Dimon called a 'liability transfer' on October 13, 2026, and it joined Anthropic's Glasswing and an Nvidia-led open-source agent safety effort.[1][2][3][4][5][6][7][8][9][10]
Data as of Oct 16, 2026 · JPMorgan profile updated Oct 13, 2026 · 19 sources cited on this page · 140 banks profiled on this site · Markdown
What frameworks and controls does JPMorgan Chase use for AI?
| Item | What it says | Who | Date |
|---|---|---|---|
| Safeguards by capability | Align safeguards to capability and risk; agents with the 'lethal trifecta' of untrusted inputs, sensitive data and authority to act require continuous enforcement | JPMorganChase technology blog | March 23, 2026[1] |
| Fine-grained agent guardrails | 'The most important thing about building agents is making sure that you have the proper risk framework in place'; guardrails 'at a very fine grain level' | Lori Beer, global CIO | October 1, 2026[2] |
| Enforceable controls for payment agents | 'A prompt expresses intent; it does not enforce authority'; an independent policy engine checks agent, principal and mandate, including pending transactions, before execution | Zack Anderson, CDAO, Payments and Global Banking | September 29, 2026[3] |
| Fence guardrail framework | Synthetic-data testing of LLM use cases for hallucinations, topic drift and prompt injection | JPMorganChase | April 2, 2026[4] |
| Trustworthy AI Center of Excellence | Firmwide center led by AI Research, with an Advisory Group from across the firm, researching explainability, fairness, robustness, security and privacy of AI, including ML models | JPMorganChase AI Research | Retrieved October 2026[5] |
| AI trust, transparency and safety | A research focus area: clear, understandable and recourse-based AI-driven decisions; includes the AlgoCRYPT Center of Excellence | JPMorganChase AI Research | Retrieved October 2026[11] |
| Secure by design | AI Threat Modeling Co-Pilot gives architects and developers threat insights and control guidance early in development | JPMorganChase | October 3, 2025[12] |
What has JPMorgan Chase disclosed or said about AI risks?
| Item | What it says | Who | Date |
|---|---|---|---|
| 10-K AI risk factor | AI system failures, biased or inaccurate output, agentic systems acting on sensitive data, AI-enabled cyber threats, evolving AI law and AI-driven workforce displacement | 2025 Form 10-K | February 13, 2026[6] |
| Cyber | AI 'will almost surely make cyber risk worse'; Dimon calls cyber the firm's largest risk | Dimon | April 2026[13][14] |
| Frontier models | Glasswing launch organization testing Claude Mythos Preview on its own systems; Dimon says frontier models first make companies more vulnerable | Anthropic; Dimon | April 7-14, 2026[9][15] |
| Outside agents | Third parties may not log in with customers' credentials; outside agents can read some data through open-banking providers but cannot move money | Bank spokeswoman; Barnum | October 13, 2026[7][16] |
| Liability transfer | 'You don't want rogue agents, and there could be thousands of them'; customers should understand the 'liability transfer'; the bank is 'not against agents or for agents' | Dimon | October 13, 2026[8][17] |
| Industry agent safety | Joins a shared, open-source safety effort for AI agents led by Nvidia | JPMorganChase | September 28, 2026[10] |
| Jurisdictional model access | Anthropic's Claude removed from the approved model list for Hong Kong staff | Reported by the Financial Times | June 18, 2026[18] |
What has JPMorgan Chase not disclosed about AI governance?
The bank publishes its engineering approach to agents and guardrails but not a standalone responsible-AI principles document, a model inventory count, or bias-testing results for credit and fraud models. Its 10-K places machine-learning models inside model risk management but gives no AI-specific governance metrics.[1][6]
There is no timeline for letting outside agents act in customer accounts; Barnum said the bank is working on secure ways to enable them, while it pilots its own Smart Cash tool.[7][8]
What does the JPMorgan profile record on this topic?
The rows below are the entries of the JPMorgan AI strategy profile that bear on this page, as of Oct 13, 2026; the profile holds the full record of 40 dated moves, 13 use cases and 67 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| Frontier-model vulnerability testing | Testing Mythos to find vulnerabilities in its own systems as a Glasswing partner. | Pilot | [15] |
| AI in model risk management and the 10-K risk factor | Models 'including those based upon machine learning or artificial intelligence techniques' sit inside estimations and model risk management; the 10-K names generative and agentic AI failures, bias, transparency and evolving AI law as risks. | In production | [6] |
| Blocking outside AI agents that use customer credentials | The bank does not let third parties log in with customers' credentials, so outside agents can at most read data through open-banking connectors and cannot move money; no timeline for direct agent access. | In production | [7][16] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Oct 13, 2026 | Q3 2026 call: outside AI agents stay blocked; AI spend put near $700 million | Dimon says the bank will not let outside agents operate inside accounts and warns of a 'liability transfer' to customers who use them; the Smart Cash pilot is under way; Barnum puts annual AI spend nearer $700 million than the $2 billion cited a year earlier, calls that figure stale and says AI benefits are not yet visible in results. | [7][8][16] |
| Sep 28, 2026 | Joins Nvidia-led push for open-source AI agent safety tools | JPMorgan signs on to a shared, open-source safety effort for AI agents led by Nvidia, as the industry works out how to give agents more autonomy without losing control. | [10] |
| Jun 18, 2026 | Anthropic access cut for Hong Kong staff | Claude models removed from the approved list in Hong Kong, following a similar move by Goldman Sachs. | [18] |
| Apr 14, 2026 | Q1 2026 call: the 'AI cash tool' is an experiment | Barnum says the consumer AI cash tool from Dimon's letter is 'kind of not even live yet' and aimed at a very small subset of clients; Dimon calls cyber the firm's largest risk and says AI 'has made it worse'. | [14] |
| Apr 14, 2026 | Testing Anthropic's Mythos; Dimon warns on cyber | JPMorgan is a named Glasswing launch partner; Dimon says frontier models first make companies more vulnerable. | [15][19] |
| Apr 7, 2026 | Named one of 12 Project Glasswing launch organisations | Anthropic's announcement lists JPMorganChase beside AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks as users of Claude Mythos Preview for defensive security work. | [9] |
| Apr 6, 2026 | Dimon's shareholder letter: AI touches 'virtually every function' | The letter says AI is not a speculative bubble, will definitely eliminate some jobs while enhancing others, and will almost surely make cyber risk worse; it also previews 'Smart Cash', which moves client cash between accounts for yield. | [13] |
| Apr 2, 2026 | Fence guardrail framework | A data-driven framework that tests LLM use cases for hallucination, topic drift and prompt injection using synthetic data. | [4] |
| Mar 23, 2026 | Engineering post on securing agents | The bank publishes its approach: align safeguards to capability and risk; agents combining untrusted inputs, sensitive data and authority to act require continuous enforcement. | [1] |
| Feb 13, 2026 | 2025 10-K adds a generative and agentic AI risk factor | The annual report lists AI system failures, biased or inaccurate output, agentic systems acting on sensitive data, AI-enabled cyber threats, evolving AI law and AI-driven workforce displacement as risks. | [6] |
How does JPMorgan Chase control AI agents?
By scaling safeguards to capability and risk. Read-only, confined agents get lighter guardrails; agents that process untrusted input, touch sensitive data and can act need continuous enforcement. In payments, the bank argues for an independent policy engine that checks each agent's mandate before any action executes, rather than relying on prompts.
What does JPMorgan's 10-K say about AI?
The 2025 Form 10-K, filed February 13, 2026, lists AI system failures, biased or inaccurate output, agentic systems acting on sensitive data, AI-enabled cyber threats, evolving AI law and AI-driven workforce displacement as risks, and places machine-learning models within its model risk management.
Why does JPMorgan block outside AI agents?
Because, a spokeswoman said, it is not safe for third parties to log in with customers' banking credentials. On October 13, 2026 Dimon warned of 'rogue agents' and a 'liability transfer' to customers who let agents act for them; outside agents can still read some data through open-banking providers but cannot move money.
Does JPMorgan Chase research explainable AI?
Yes. Its AI Research group leads a firmwide Trustworthy AI Center of Excellence, with an advisory group from across the firm, that researches explainability, fairness, robustness, security and privacy of AI and publishes at conferences such as ICLR and ICML.
- Securing the next generation of AI agents
- The Most Powerful Women in Banking, #5, Lori Beer, JPMorganChase
- Why AI Payment Agents Need Enforceable Controls, Not Prompts
- Strengthening LLM guardrails with synthetic data generation
- Trustworthy AI Center of Excellence
- JPMorgan Chase & Co. Annual Report on Form 10-K for the year ended December 31, 2025
- JPMorgan Chase blocks outside AI agents like Muse, Instinct from directly accessing accounts
- JPMorgan CEO Dimon warns customers of 'liability transfer' with AI
- Project Glasswing: Securing critical software for the AI era
- JPMorgan Stock Slips 1% as Bank Agents Get an Open Safety Layer
- Artificial Intelligence Research
- Revolutionizing Threat Modeling with AI: The Threat Modeling Co-Pilot
- Jamie Dimon's Letter to Shareholders, Annual Report 2025
- JPMorgan Chase & Co. Q1 Earnings Call Highlights
- Jamie Dimon says Anthropic's Mythos reveals 'a lot more vulnerabilities' for cyberattacks
- JPMorgan warns of risks to consumers who rely on AI agents
- Jamie Dimon Says JPMorgan Isn't Using External AI Agents
- JPMorgan Chase cuts off Anthropic access for its Hong Kong staff
- Banks in Asia brace for complex cyber threats from frontier AI
When JPMorgan moves on AI, you'll read it here first.
when one of the banks we cover moves on AI, it is in the next morning's brief · six sourced stories · 7 am ET · free
plus every tracker, bank and agent page update, the morning after · leave any morning