Claude Mythos and Banks: Protecting Legacy Banking Systems When AI Finds Vulnerabilities Faster
What Claude Mythos means for banks: what regulators expect, why legacy and end-of-life systems are most exposed, and how to govern AI agent access.

Claude Mythos matters to banks because it can find and exploit software vulnerabilities faster and at lower cost than skilled people, and banks run many older systems where unpatched weaknesses build up. Anthropic has kept Claude Mythos Preview in a gated research preview through Project Glasswing, whose launch partners include JPMorganChase, while UK authorities have told financial firms to find and fix vulnerabilities faster and US officials have raised it directly with the largest banks.
For a bank or credit union, the practical answer is not a new kind of defense. It is the basics done faster and more completely: patching, access management, network security and logging, with particular attention to end-of-life systems and to every new path that AI tools open into core systems.
Where engini fits: engini is not a vulnerability scanner and does not patch systems. It is the governed path through which AI agents and workflows reach a bank's systems, including databases and file drops behind the firewall. Its on-prem agent opens no inbound ports, toolsets limit which tools and which connection an agent may use, engini acts with the permissions of the connected account, and an audit log records each call, so new AI access does not become a new attack surface. Start with the engini docs index.
Key Takeaways
- Claude Mythos Preview is a general-purpose Anthropic model announced on April 7, 2026. Anthropic says it has identified thousands of zero-day vulnerabilities and offers it only as a gated research preview through Project Glasswing.
- The UK AI Security Institute found that Mythos Preview succeeds on 73% of expert-level capture-the-flag tasks and was the first model to complete its 32-step corporate network attack simulation, in 3 of 10 attempts.
- The Bank of England, the FCA and HM Treasury said on May 15, 2026 that the cyber capabilities of current frontier AI models already exceed what a skilled practitioner could achieve, and named end-of-life and out-of-support systems as an area of increased exposure.
- The OCC's Spring 2026 risk report says multifactor authentication and timely patch management help mitigate AI-enabled cyber risks for banks.
- Every new connection an AI agent uses into a core or legacy system should be scoped to what it needs, avoid inbound firewall openings where possible, and be logged.
What is Claude Mythos?
Claude Mythos Preview is a general-purpose frontier model from Anthropic that is especially strong at finding and fixing software vulnerabilities. Anthropic describes it as its most capable model yet for coding and agentic tasks, and says its strength in cybersecurity follows from that: a model that can deeply understand and modify complex software can also find its weaknesses (Anthropic: Project Glasswing).
Anthropic has not released Mythos Preview to the public. It is available as a gated research preview to participants in Project Glasswing, an initiative announced on April 7, 2026 "to secure the world's most critical software for the AI era." The launch partners are Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, and Anthropic says it extended access to over 40 additional organizations that build or maintain critical software. Participants reach the model through the Claude API, Amazon Bedrock, Google Cloud's Vertex AI and Microsoft Foundry.
According to Anthropic, Mythos Preview "has already identified thousands of zero-day vulnerabilities across critical infrastructure." JPMorganChase's chief information security officer said on the same page that the bank will "take a rigorous, independent approach" to deciding how to proceed.
Why does Claude Mythos worry banks?
Banks worry because a model that lowers the cost of finding weaknesses helps attackers as much as defenders, and the financial system depends on software that is hard to change quickly. In April 2026, finance ministers, central bankers and bank chief executives raised the issue publicly (BBC News, April 17, 2026):
- Bank of England Governor Andrew Bailey told the BBC that AI could make it "easier to detect existing vulnerabilities in sort of core IT systems" for criminals to exploit.
- Canada's Finance Minister François-Philippe Champagne said Mythos had been discussed extensively at the IMF meeting in Washington and called it "the unknown, unknown."
- Barclays chief executive CS Venkatakrishnan said: "It's serious enough that people have to worry."
- The BBC reported that the US Treasury had raised the issue with major banks and encouraged them to test their systems before any public release.
In a May 2026 speech, Federal Reserve Vice Chair for Supervision Michelle Bowman said the Treasury Secretary and the Fed Chair had convened the largest banks to discuss Mythos, and that "banks of all sizes have expressed concern about access to the Mythos model" (Federal Reserve: Bowman, Artificial Intelligence in the Financial System). She also noted the other side: the same capability lets firms find and fix their own vulnerabilities.
What did the AI Security Institute find?
The UK AI Security Institute (AISI) found that Mythos Preview is a step up over earlier frontier models at cyber tasks and can attack small, weakly defended networks on its own once it has network access (AISI: Our evaluation of Claude Mythos Preview's cyber capabilities, April 13, 2026).
AISI is careful about the limits. Its ranges lack active defenders and defensive tooling, and the model is not penalized for setting off alerts, so AISI "cannot say for sure" whether Mythos Preview could attack well-defended systems. Its advice to organizations is plain: get the cybersecurity basics right, which AISI lists as regular security updates, access controls, security configuration and "comprehensive logging."
Why are legacy banking systems more exposed?
Legacy banking systems are more exposed because they are harder to patch, often run past vendor support, and carry decades of interfaces that were built before anyone expected an attacker to read code at machine speed. The UK authorities name this directly: investment and resourcing should reflect "increased exposure from end-of-life systems or those out of vendor support" (Bank of England, FCA and HM Treasury joint statement, May 15, 2026).
In practice, the risk in an older estate tends to sit in a few places:
These are views on where to look first, not findings about any one institution. Your own inventory decides the order.
What are regulators asking banks to do?
Regulators are asking banks to apply existing cyber and operational resilience expectations faster and at larger scale, not to follow a new rulebook. The UK statement says it is "not intended to introduce new expectations" but sets out what firms should do across five areas.
In the US, the OCC's Spring 2026 Semiannual Risk Perspective says AI enables "rapid vulnerability discovery and exploitation" and that "multifactor authentication and timely patch management" help mitigate AI-enabled cyber risks (OCC Semiannual Risk Perspective, Spring 2026).
Does connecting AI agents to legacy systems add risk?
Yes, if the agent gets broad access: every connection an AI agent uses into a core system is one more path an attacker can try, and one more set of credentials to protect. Banks are also adopting AI agents for their own work, such as reconciliation, exception handling and reporting, so the question is how those agents reach the core, not whether they will.
Four rules keep agent access from widening the attack surface:
- Scope the tools. An agent should see only the actions it needs. If it only reads balances, it should not have a write action available at all.
- Limit the account. Connect the agent through a database or system account with the narrowest rights that work. Read-only accounts keep read-only agents read-only, whatever the prompt says.
- Open nothing inbound. Prefer integrations that connect outward from your network over ones that need an inbound firewall rule.
- Log every call. Each action an agent takes should land in the same audit trail and SIEM as the rest of your estate.
Prompt instructions are not controls. The PocketOS incident, where an AI coding agent deleted a production database and its backups despite a written rule not to, shows why limits have to be enforced outside the model.
How engini helps banks govern AI access to legacy systems
engini gives AI agents and workflows a governed way into the systems a bank already runs, so you can use AI on legacy data without opening new inbound paths or handing out broad credentials. It does not find or fix vulnerabilities; that remains the job of your security tools and teams.
- No inbound ports. The on-prem agent is a small service on your own server. It opens every connection itself, out to engini over HTTPS on port 443, so you open no inbound ports or firewall rules. It reaches on-premises databases (SQL Server, Oracle, IBM DB2, MySQL, PostgreSQL and ODBC sources), file drops in local folders and network shares, and internal REST and HTTP applications. IBM AS/400 systems, the classic core in many older financial institutions, are reached the same way.
- Toolsets limit what an agent can call. A toolset lists the tools an agent may use and which connection each one uses, for example read-only. engini enforces it on the server: the agent discovers only those tools, and a call outside the toolset is refused. Toolsets scope calls through the CLI, SDK and REST API; an MCP server exposes only the connections and tools chosen when it is set up.
- The agent acts as the account you connect. engini acts with the permissions of the connection it uses, so a read-only database account keeps an agent read-only. Every run is recorded under the user who triggered it.
- Credentials stay out of prompts. Your agent never needs your credentials: engini stores them and adds them when it runs each call, so keys stay out of prompts and tool inputs.
- Everything is logged. The audit log can be exported and streamed to your SIEM, and each call through the on-prem agent writes a history record.
- Platform controls. Single sign-on, two-factor authentication, IP restrictions, session timeout, roles including custom roles, and separate development, staging and production environments. engini holds SOC 2 Type II and ISO 27001; details are on the security page.
What engini does not do: it does not scan for or patch vulnerabilities, detect or mask sensitive data, require a human approval before an agent acts, or discover agents that do not connect through engini. Plan those controls elsewhere.
When you do not need engini: if no AI agent or automation touches your core and legacy systems, your priority is vulnerability management, patching and segmentation, and engini adds nothing there. engini fits when you want AI agents and workflows to work with core, ledger and file-based systems under the controls above. More on that pattern: MCP for finance systems and legacy cores, SOAP-to-REST orchestration for legacy banking cores, and engini for financial services.
Frequently Asked Questions
What is Claude Mythos in banking?
Claude Mythos is an Anthropic AI model that finds software vulnerabilities, and banks treat it as a cyber risk to older, hard-to-patch systems. Anthropic offers it only to Project Glasswing participants, including JPMorganChase, for defensive security work.
Can banks get access to Claude Mythos?
Access is limited to Project Glasswing participants during the research preview. JPMorganChase is a launch partner and Anthropic says over 40 more organizations have access; the Fed's Michelle Bowman said banks of all sizes have raised concerns about access.
What did the Bank of England say about Claude Mythos?
Governor Andrew Bailey said the risk had to be taken very seriously because AI could make vulnerabilities in core IT systems easier to find. On May 15, 2026 the Bank, the FCA and HM Treasury told firms to manage vulnerabilities faster and reduce their attack surface.
Does Claude Mythos make legacy core banking systems unsafe?
Not by itself, but it lowers the cost of finding and exploiting weaknesses that already exist. The UK AI Security Institute found it can exploit systems with weak security posture, and could not say whether it can attack well-defended systems.
How should banks control AI agents that access core systems?
Give each agent only the tools it needs, connect it through a least-privilege account, avoid inbound firewall openings, and log every call into your SIEM. Instructions in a prompt are not a control.
Sources
- Anthropic: Project Glasswing and Claude Mythos Preview
- UK AI Security Institute: Our evaluation of Claude Mythos Preview's cyber capabilities
- Bank of England, FCA and HM Treasury: Joint statement on frontier AI models and cyber resilience
- Federal Reserve: Vice Chair for Supervision Bowman, Artificial Intelligence in the Financial System
- OCC: Semiannual Risk Perspective, Spring 2026
- BBC News: Finance ministers and bankers raise serious concerns about Mythos AI model
- engini docs: On Prem Agent and Scope an agent with a toolset