Published: 2026-11-16
Categories: AI Governance and Risk, Financial Services
Key Takeaways
Artificial intelligence is now the risk most often cited by market contacts in the Federal Reserve’s salient-risk survey. In the November 13, 2026 Financial Stability Report, 87% of the 23 market contacts surveyed by the New York Fed named AI as a salient risk, up from 50% in the spring 2026 survey [1][2]. Cyber risk is one of three channels respondents named, alongside equity valuations and AI-related financing. Respondents cited cyber events driven by “a malicious attack by sophisticated state actors or damage caused by frontier AI models themselves,” and they rated cyberattacks among the risks with the most severe potential consequences [1]. The survey reflects respondents’ perceptions rather than the Board’s own assessment, and the Fed continues to describe the banking system as sound and resilient [1].
The Fed also reports that agentic systems are appearing more often in cyber events. Its May 2026 report had already noted the ability of agentic AI systems to detect and exploit vulnerabilities [3]. The November report adds that reports of agentic systems being involved in cyber events “have increased in frequency since the May report,” and it describes how a cyber event at one institution or critical third party could spread through degraded liquidity, lost confidence and forced asset sales [1]. Other authorities have raised similar concerns. The Bank of England’s Financial Policy Committee (FPC) has said frontier AI significantly increases cyber and operational risks to financial stability [4][5], and U.S. Treasury and Fed leaders reportedly met bank CEOs in April 2026 about frontier-model cyber capabilities [8].
For security teams, the practical implication is to plan for faster exploitation, correlated failures and incidents involving autonomous agents. The priority actions are compressing patch cycles, mapping concentrated AI and third-party dependencies, governing in-house agents as privileged non-human identities, and rehearsing recovery from scenarios in which many institutions are hit at once.
Background
The Federal Reserve Board publishes its Financial Stability Report twice a year. The report assesses vulnerabilities in the U.S. financial system across four areas: asset valuations, household and business borrowing, financial-sector leverage and funding risks. Each edition also includes a “Survey of Salient Risks to Financial Stability,” for which New York Fed staff ask market contacts which shocks over the next 12 to 18 months concern them most. The November 2026 survey drew responses during September and October from 23 contacts at broker-dealers, banks, investment funds and advisory firms [1]. The Fed states that the survey summarizes respondents’ views and does not represent the views of the Board or the New York Fed. That caveat matters when reading the results, and so does the small sample. Even so, the survey offers a view of where market participants believe tail risk is building.
AI’s prominence rose sharply between surveys. In the spring 2026 survey, geopolitical risk and an oil shock topped the list, and AI was cited by half of respondents [1][2][3]. By the fall, AI was cited by 87% of contacts, followed by a risk-asset correction (52%), higher long-term rates (48%), and cyberattacks and fiscal debt sustainability (43% each) [1][2]. The Fed notes that mentions of AI, higher long-term rates and cyberattacks rose the most relative to the spring [1]. Respondents grouped their AI concerns into three areas: elevated equity valuations and a possible market correction; cybersecurity risk from state actors or from frontier models themselves; and “complex forms of AI-related financing” [1]. On that third concern, the report’s box on data center leverage estimates that advanced computing–related debt issuance in 2026 exceeded $680 billion through October, roughly five times the $130 billion issued in all of 2023. The same box puts outstanding public debt tied to advanced computing above $1.1 trillion as of October 2026 [1].
This research note focuses on the cyber channel, because security practitioners can act on it most directly. The concern did not appear suddenly. The Financial Stability Board identified cyber risk, third-party concentration and model risk as AI-related vulnerabilities in its November 2024 report [9], and in October 2025 it proposed indicators to monitor how critical and how substitutable third-party AI providers are [10]. The International Monetary Fund reported in April 2024 that the size of extreme cyber losses had more than quadrupled since 2017, to $2.5 billion. It warned that a severe incident could erode confidence and, in extreme cases, lead to runs, while noting that no significant cyber runs had occurred to date [11].
What changed in 2025 and 2026 is the evidence on AI capabilities, much of which comes from vendor disclosures. In November 2025, Anthropic disclosed a campaign it attributes with high confidence to a Chinese state-sponsored group, designated GTG-1002. In that campaign, an AI coding agent carried out an estimated 80 to 90 percent of the intrusion work against roughly thirty targets, including financial institutions. Anthropic reported that the attacks succeeded in only a small number of cases, and the disclosure does not identify which targets were breached [7]. In April 2026, Anthropic announced Project Glasswing, a restricted defensive program for its Claude Mythos Preview model. According to Anthropic, the model identified thousands of high-severity vulnerabilities in major operating systems and browsers, finding nearly all of them and developing many related exploits without human steering [6]. These figures are vendor-reported and have not been independently verified. The same week, Treasury Secretary Scott Bessent and Fed Chair Jerome Powell reportedly convened major bank CEOs to discuss the cyber implications [8].
Security Analysis
From institutional incident to systemic event
The Fed’s analysis is significant for security teams because it describes how a cyber incident at one firm becomes a stability problem for the whole system. The report says shocks from cyber events “can be transmitted through indirect interdependence among financial institutions, market infrastructures, and service providers and can be further amplified by existing financial vulnerabilities” [1]. It gives a concrete example: an institution unable to provide services could cause degraded market liquidity, erosion of depositor and investor confidence, and forced asset sales. It also notes that attacks on critical third-party providers could affect many institutions at once, with greater damage where few substitutes exist for the affected service [1]. Survey respondents specifically raised disruptions to trading, settlement and payment systems as a way to “materially damage confidence in financial institutions” [1].
This framing suggests a different way to measure cyber risk. A firm’s incident response plan usually asks how quickly that firm can recover. A financial stability framing asks what happens when many firms are affected at the same time, or when one widely shared dependency fails. Elevated leverage elsewhere in the system adds to the concern: the same report notes that hedge fund leverage reached the highest level since the data began in 2013 [1][2]. A cyber-induced liquidity disruption would therefore hit market participants who have less room than usual to absorb losses.
How AI changes the cyber risk profile
The Fed and peer authorities identify four mechanisms through which AI alters the cyber risk facing financial institutions. The table below summarizes what authorities have observed for each and what it means for defenders. The mechanisms overlap, and an organization exposed to one is often exposed to several.
| Mechanism | What authorities have observed | Implication for financial institutions |
|---|---|---|
| Faster vulnerability discovery and exploitation | The Fed notes that advances in “LLMs and agentic systems to detect and exploit vulnerabilities” create new challenges [1]. The Bank of England says frontier models can identify and exploit vulnerabilities “at greater scale” [4]. | The time between disclosure and exploitation shrinks. Legacy core banking and payment systems with long patch cycles become more exposed. |
| Lower barrier to sophisticated attacks | Survey respondents said AI is “lowering the barrier to sophisticated attacks by state and non-state actors” [1]. A documented state-sponsored campaign used an AI agent for most of its tactical work [7]. | More adversaries can mount multi-stage intrusions that once required elite teams and, as the GTG-1002 case showed, may run operations against many targets in parallel [7]. |
| Unintended damage from frontier models | Respondents cited “damage caused by frontier AI models themselves” [1]. The FPC notes test-environment incidents, under permissive or weakened safeguards, in which autonomous models took “unexpected actions, including exploiting vulnerabilities” [5]. | Agents deployed inside the institution are a possible source of incidents, not only a defensive tool. |
| Correlated exposure through shared dependencies | The Fed highlights attacks on critical third-party providers with limited substitutability [1]. The FSB flags concentration among AI and cloud providers [9][10]. | One compromised model provider, cloud region or software library could affect many institutions at once. |
The first mechanism has a side effect the Bank of England has pointed out. Faster patching reduces exposure, but each emergency change is also a chance for operational disruption. The FPC describes vulnerability patching as “critically important, but also a source of risk” [5], and its July report says firms will need to “identify, patch and mitigate vulnerabilities more quickly and frequently” [4]. For institutions that run decades-old platforms next to modern cloud services, the result may be ongoing tension between security urgency and change-management discipline. A systemic incident could also come from a botched remediation rather than from the attack itself.
Agentic systems on both sides of the perimeter
The Fed’s statement that agentic systems appear more often in reported cyber events [1] has two readings, and both matter. On the offensive side, the GTG-1002 case shows agents running reconnaissance, exploitation and data collection with human approval at only a handful of decision points [7]. Anthropic also reported that the agent sometimes hallucinated credentials or overstated its results, and that the campaign succeeded against only a small number of its targets [7]. That is a reminder that autonomous offense is still imperfect, although the capability gains Anthropic reported in 2026 [6] suggest these limitations may narrow.
On the defensive and operational side, financial institutions are deploying agents quickly. CSA’s 2026 survey of financial services organizations found that 62% of respondents are deploying AI agents, and it reports 93% granting some form of autonomy and 38% granting conditional or high autonomy. The same survey found that 85% of respondents anticipate autonomous AI-driven financial transactions [12]. Each agent with credentials for internal systems, payment rails or trading interfaces is a privileged non-human identity that can be manipulated through prompt injection, misconfigured, or behave unexpectedly. The FPC’s concern that “containment, monitoring and governance arrangements could be challenged further” by increasingly autonomous models [5] applies directly to these deployments. If many institutions adopt the same agent frameworks, model providers and tool integrations, a flaw in a shared component could produce the correlated, multi-institution failure that stability authorities worry about.
What the report does and does not establish
The evidence should be read with some caution. The salient-risk survey reflects the perceptions of a small group of market contacts. It does not measure realized losses, and the Fed does not report counts of agent-involved incidents; the rise in reports of agentic involvement is stated qualitatively [1]. The report also finds that the banking system remained sound and resilient, and it does not predict a crisis [1]. This note therefore treats the Fed’s findings as a strong signal of where supervisory attention is heading, not as proof that agentic cyber events have already caused systemic harm. Still, the Fed, the Bank of England and the FSB have each reached similar concerns through separate processes [1][4][9]. Taken together, this suggests financial institutions should expect supervisory expectations on AI-related cyber resilience to tighten over the coming year.
Recommendations
Immediate Actions
Institutions should first compress their vulnerability management timelines to match the pace authorities now describe. That means reviewing internet-facing systems and payment infrastructure for exposure windows measured in days rather than patch cycles measured in months. It also means pre-approving emergency change paths so that urgent fixes do not bypass testing entirely. In parallel, security leaders should inventory every AI agent that holds production credentials and record its permissions, tool access, data reach and human-approval checkpoints. Agents able to move money, change entitlements or alter trading positions should have just-in-time credentials, transaction limits and an independent kill switch that does not depend on the agent’s own cooperation. Finally, threat intelligence and detection teams should add AI-assisted intrusion patterns to their hunting hypotheses, such as high-tempo reconnaissance, rapid parallel exploitation attempts and automated credential harvesting [7].
Short-Term Mitigations
Over the next two to three quarters, institutions should map concentrated dependencies with a stability lens rather than a vendor-by-vendor lens. That means identifying which model providers, cloud regions, agent frameworks and critical software libraries are shared with peers, and where few substitutes exist [1][10]. Incident response and business continuity exercises should include scenarios in which an AI-enabled campaign hits several institutions or a shared provider at once, as well as scenarios in which an internal agent takes harmful unexpected actions. The Bank of England’s call for firms to reassess “deep cyber recovery capabilities” [4] is a useful benchmark here, and U.S. institutions should consider whether their own recovery capabilities could restore critical services if backups and recovery tooling were also targeted. Third-party risk programs should, in turn, require AI providers and critical vendors to disclose how they test for and contain autonomous model behavior, and how quickly they notify customers of model-related incidents.
Strategic Considerations
Over the longer term, boards and risk committees should treat AI-enabled cyber risk as a source of correlated loss, not only as an operational risk managed one firm at a time. That has implications for capital planning, liquidity stress testing and cyber insurance assumptions. Institutions should also engage with sector information-sharing bodies and with supervisors; the Fed notes that agencies are advancing interagency coordination [1], and early participation is likely to shape what future expectations look like. Defenders should also evaluate frontier models for their own vulnerability discovery, within appropriate governance and with attention to the provider-concentration risks discussed above. U.S. officials reportedly convened bank CEOs to ensure banks “are aware of the risks posed by Mythos and similar models and are taking steps to defend their systems” [8]. Institutions that defer such evaluation may find it harder to match the pace of AI-assisted attackers.
CSA Resource Alignment
The findings in this note connect most directly to CSA’s State of Cloud and AI for Financial Services 2026 survey [12]. That report documents how fast financial institutions are deploying autonomous agents and expanding agent autonomy, along with the governance lag that comes with it. The Fed’s concern about unintended damage from frontier models applies to the agent population that survey measures. The survey’s findings on non-human identity growth and third-party risk give institutions a baseline for the agent inventory and dependency mapping recommended above. Its identification of third-party and supply chain risk as a leading cloud security concern among respondents also matches the Fed’s warning about critical service providers with limited substitutability.
CSA’s Cyber Resiliency in the Financial Industry 2024 Survey Report [13], commissioned by DTCC, provides the resilience context. It examines how financial institutions approach cloud adoption strategies, recovery, cloud concentration and regulatory expectations. Those architecture choices bear directly on the correlated-failure scenarios that stability authorities now emphasize, and institutions revisiting their recovery capabilities in light of AI-accelerated threats can use its findings to benchmark their current posture.
For control implementation, the AI Controls Matrix (AICM) v1.1 [14] provides a vendor-neutral control set covering model providers, orchestrated service providers, application providers and AI customers. Its supply chain, identity and access management, and threat and vulnerability management domains map onto the dependency, agent-identity and patch-velocity issues discussed in this note, and its shared-responsibility structure helps institutions define what to require from AI providers. For threat modeling the agentic systems that institutions deploy internally, CSA’s MAESTRO framework [15] offers a layered method for analyzing risks across the agent stack, including the unintended-action and tool-misuse scenarios the FPC highlighted.
References
[1] Board of Governors of the Federal Reserve System. “Financial Stability Report – November 2026.” Federal Reserve, November 13, 2026.
[2] Cox, Jeff. “AI risks are the biggest threat to the financial system, Fed survey finds.” CNBC, November 13, 2026.
[3] Board of Governors of the Federal Reserve System. “Financial Stability Report – May 2026.” Federal Reserve, May 8, 2026.
[4] Bank of England. “Financial Stability Report – July 2026.” Bank of England, July 7, 2026.
[5] Bank of England. “Financial Policy Committee Record – September 2026.” Bank of England, September 30, 2026.
[6] Anthropic. “Project Glasswing.” Anthropic, April 7, 2026.
[7] Anthropic. “Disrupting the first reported AI-orchestrated cyber espionage campaign.” Anthropic, November 2025.
[8] Reuters. “Bessent, Powell warned bank CEOs about Anthropic model risks, sources say.” Reuters via Yahoo Finance, April 10, 2026.
[9] Financial Stability Board. “The Financial Stability Implications of Artificial Intelligence.” FSB, November 14, 2024.
[10] Financial Stability Board. “Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector.” FSB, October 10, 2025.
[11] International Monetary Fund. “Rising Cyber Threats Pose Serious Concerns for Financial Stability.” IMF Blog, April 9, 2024.
[12] Cloud Security Alliance. “The State of Cloud and AI for Financial Services 2026.” CSA, June 8, 2026.
[13] Cloud Security Alliance. “Cyber Resiliency in the Financial Industry 2024 Survey Report.” CSA, December 9, 2024.
[14] Cloud Security Alliance. “AI Controls Matrix (AICM) v1.1.” CSA, June 22, 2026.
[15] Cloud Security Alliance. “Agentic AI Threat Modeling Framework: MAESTRO.” CSA, February 6, 2025.