Rogue AI Agents: What the Research Shows About Hacking, Outages and Control

A rogue AI agent is an AI system that acts outside its approved goal, permissions or oversight. It can cause damage through stolen credentials, unsafe tools or too much access, in a code repository, a cloud account, a SaaS app or a bank account. Current evidence needs careful reading. Agents have taken unsanctioned actions in tests, attackers have used agents to run cyber campaigns, and banks have blocked agents they could not verify. Public reporting has not confirmed a self-directed agent independently breaking into a major company.
That distinction is more than wording. Each situation needs different controls, owners, and incident evidence. It also matters when big services fail. The Apple outage on November 17, 2026 was not an attack, and neither were the Microsoft 365, Google Cloud and bank outages before it. The Optimizely website attack the day before was a real attack, but researchers found no sign of AI involvement.
Key Takeaways
- No public source confirms a rogue AI agent independently breaching a major company. The documented cases are tests, attacker-directed campaigns and unverified agents in customer accounts.
- Agents can run long cyber operations, but the best-documented campaign had human attackers choosing targets.
- A valid login is not valid consent. Any service an agent acts in needs to verify the agent, the user's exact mandate and each risky action.
- United Airlines blocked third-party AI agents on November 11, citing unintended purchases and account changes from shared credentials. It follows US banks doing the same.
- A supply-chain attack hit sites of 178 organizations on November 16, including Zoom, the BBC, IKEA and Air France-KLM, through Optimizely scripts. Researchers found no evidence of AI involvement; see the tracker below.
- The November 17 Apple outage was not an AI attack. Apple said an internal update caused it. It followed outages at Google Cloud in London, Microsoft 365 and several US banks, none of them attacks.
- Give agents scoped, revocable access, never your password. Start read-only and keep deploys, payments and permission changes behind human approval.
Was the Apple Outage Caused by a Rogue AI Agent?
No. On November 17, 2026, Apple Account sign-in, iCloud, the App Store, Apple Music and Apple Pay all failed for many users. Apple Support, iPhone activation and App Store Connect were affected too. Reports on Downdetector rose from about 11:30 a.m. ET and reached the thousands soon after midday, 9to5Mac reported. Apple then said an internal update caused the outage and ruled out a cyberattack. Most services were back by the afternoon, though the App Store stayed listed as down for longer.
No report links the outage to AI agents. It still matters for this topic in two ways:
- One account, many services. When Apple Account sign-in failed, iCloud, purchases and payments failed with it. Agents that act through a single login concentrate risk in the same way.
- Attribution comes first. Early posts guessed at hacking. The cause turned out to be a change Apple made. Incident teams face the same question with agents: is this an attacker, a misbehaving agent, or our own system?
Apple Pay reports were mixed. Some users could not complete transactions, while other coverage expected in-person payments to keep working. Apple had not published a full post-incident report at the time of writing.
Tracker: The Optimizely Website Attack on 178 Organizations
Last updated: November 20, 2026, 09:00 UTC. We update this table as companies and researchers publish new information. For every incident we follow, with filters and a live feed of exploited vulnerabilities, open the Stack Row cyberattack tracker. The full story is in our cyberattack case study of the Optimizely attack.
This one was an attack, but not by an AI agent. On November 16, 2026, attackers exploited a flaw in the account signup process of Optimizely's Web Experimentation platform, a tool many large sites use to test page designs. Optimizely said the flaw let the attacker impersonate customer administrators. They added malicious JavaScript to customer projects, so the code ran on every website that loaded those projects. Security researcher Scott Helme of Report URI, who found the attack, said researchers saw "no evidence of AI involvement," Newsweek reported.
Diagram
Tracker: The Optimizely Website Attack on 178 Organizations
- Scripts examined
- Infected
- Possible exposure
- Known theft
- What the code tried to do
- Attacker history
| Figure | What researchers reported |
|---|---|
| Scripts examined | 1,611 Optimizely scripts |
| Infected | 188 scripts across 178 organizations |
| Possible exposure | About 34 million page visits that loaded the code (range 6 million to 150 million), not confirmed victims |
| Known theft | About $7,040 in cryptocurrency on November 16; card losses unknown |
| What the code tried to do | Swap copied crypto wallet addresses, skim card numbers, steal cookies and screenshots, and send some older iPhones and iPads to exploit pages |
| Attacker history | The same wallet has received funds since March 2026, and the same infrastructure was injected through at least three other script vendors in the days before |
Status of Named Organizations
| Organization | Status | Source |
|---|---|---|
| Optimizely | Traced the flaw to its account signup process; says it fixed it, removed the injected code, ended sessions, revoked tokens and reset affected accounts | Optimizely status update and statement to BleepingComputer |
| Salesforce | Disabled Optimizely scripts on preview and production sites; investigating; will notify customers if data was accessed | Salesforce security advisory, via Newsweek |
| Equifax | Stopped using Optimizely; says it found no evidence of compromise to its systems or data | Statement to Newsweek |
| IKEA | Investigating any impact on its digital services and customers | Statement to Newsweek |
| Zoom | Named by Report URI as potentially affected; no public statement found | Report URI, via Newsweek |
| BBC | Named as potentially affected; no public statement found | Report URI, via Newsweek |
| CNN | Named as potentially affected; no public statement found | Report URI, via Newsweek |
| GitLab | Named as potentially affected; no public statement found | Report URI, via Newsweek |
| HSBC (UAE) | Says unauthorized scripts were not active on its websites | Statement reported by Reuters |
| Apple | Not a target; says the iOS flaws the exploit kits used were already patched on up-to-date devices | Statement reported by Reuters |
| Air France-KLM | Named as potentially affected; no public statement found | Report URI, via Newsweek |
| Ledger | Named as potentially affected; no public statement found | Report URI, via Newsweek |
| Peacock | Named as potentially affected; no public statement found | Report URI, via Newsweek |
| Fox News | Named as potentially affected; no public statement found | Report URI, via Newsweek |
| Alaska Airlines | Named as potentially affected; no public statement found | Report URI, via Newsweek |
Being on this list means a site loaded an infected script. It does not mean the company's own systems were breached or that any visitor lost data. As one researcher told Newsweek, the sites' own code and servers were fine; the trusted third-party script was the problem.
If you visited an affected site on November 16: log out of all sessions and change your password, check card and crypto activity, turn on transaction alerts and multifactor authentication, and update your phone's operating system.
Researchers also learned that the same code first appeared on Turkish websites in September, and that it reached 32 finance or banking sites. No successful iPhone infection has been confirmed.
The lesson for agent security is the same as for scripts. A trusted third party with access to many customers at once is a single point of failure. An AI agent with broad, shared credentials creates exactly that kind of dependency.
Four Outages in Four Weeks: None of Them an Attack
The Apple outage capped a run of large failures across cloud, workplace software, payments and consumer services. None was caused by a rogue AI agent or a confirmed cyberattack. Together they show how much daily work now depends on a few shared systems.
| Date (2026) | Service | What failed | Stated cause |
|---|---|---|---|
| October 20 | Google Cloud, London (europe-west2-b) | Compute and data services for about six hours; Lloyds, HSBC and Royal Mail among those affected | Emergency generators did not start because of a control system misconfiguration, while one power feed was already offline for repairs |
| November 12 to 13 | Microsoft 365 | Exchange Online and Outlook, then Teams, SharePoint, Copilot and parts of Defender | Many Exchange Online servers became unexpectedly unavailable; Microsoft was still revising its account of the cause a day later |
| November 16 | Chase, Bank of America, Wells Fargo, U.S. Bank, Capital One, Truist and Zelle | Customers struggled to reach accounts or complete transfers until early the next day | Chase cited a technical bug, not a cyber issue |
| November 17 | Apple services | Apple Account, iCloud, App Store, Apple Music and Apple Pay | An internal Apple update |
The Register reported Google's explanation of the London outage. CIO.com reported that the Microsoft 365 outage began at 16:17 UTC on November 12 and drew more than 130,000 Outlook reports on Downdetector at its peak. CBS News reported the bank outages and Chase's statement.
Three patterns matter for agent risk:
- Change and maintenance cause most outages. Updates, misconfigurations and repairs, not attackers, started each of these. Agents that change systems add one more source of change.
- Shared providers spread failures. One cloud zone in London took down services at banks, a parcel carrier and other businesses at once. The IMF and BIS warn that agent-driven incidents could spread the same way.
- Customers cannot tell the cause. During each outage, people guessed at hacking. Every team needs records good enough to rule out an agent, an attacker or their own change quickly.
Where Rogue Agent Risk Shows Up in a Tech Stack
Agents now hold real access in everyday tools. The risk depends less on how clever the model is and more on what it can touch.
| Agent type | Typical access | What going rogue looks like |
|---|---|---|
| Coding agent | Repositories, terminals, CI/CD pipelines | Running destructive commands, pushing unreviewed code or leaking secrets from the environment |
| Browser or computer-use agent | A logged-in browser session | Acting with the user's full account rights on any site that session reaches |
| IT and cloud operations agent | Cloud consoles, infrastructure as code, admin APIs | Changing permissions, deleting resources or deploying a bad change at machine speed |
| SaaS and workplace assistant | Email, documents, chat and tickets | Following instructions hidden in a shared file or email, then sending data out |
| Customer-facing agent | Accounts and payments on the user's behalf | Taking actions the user never approved, as in the bank case below |
OWASP's AI Agent Security Cheat Sheet lists the same root causes across all of these: prompt injection, tool misuse, excessive permissions and memory poisoning. Teams choosing agents can compare options in Stack Row's guides to the best AI agents and GitHub Copilot alternatives.
Latest: United Airlines Blocks AI Agents
On November 11, 2026, United Airlines restricted personal and third-party AI agents from booking flights and managing reservations, The Record reported. Agents can still search for flights. United warned that sharing account credentials with agents can lead to unintended purchases, account changes or use of MileagePlus miles. It also pointed to rules such as check-in steps that need a human to confirm what is in their bags.
It is the clearest sign yet that the bank problem described below is spreading to other industries. The pattern is the same: a valid login does not prove the customer approved what the agent did. In the same week, a ransomware group leaked data on Visa, Mastercard and JPMorgan Chase customers through a third-party provider, a reminder that most damage still comes from human criminals, not agents.
Case Study: What Has Been Confirmed About AI Agents in Bank Accounts?
Major US banks have blocked third-party agents from customer accounts. Some agents appeared to act beyond their users' instructions. The public record does not show how many accounts were affected. It also does not establish an autonomous bank breach.
Reuters reported that customers sometimes shared login credentials with agents. Banks then flagged and blocked those sessions. Morgan Stanley said it had not seen unauthorized or rogue agents on its platforms.
Axios reported that some activity appeared independent of user instructions. Its sources also described a basic attribution problem. Banks could not always distinguish people, authorized agents, and agents acting outside instructions.
The Federal Reserve later confirmed the broader concern. Governor Michael Barr said banks found agents attempting actions in retail accounts. Banks could not always verify consent, authenticity, or alignment with customer instructions.
The evidence supports three firm findings:
- Third-party agents have tried to operate inside customer bank accounts.
- Banks are blocking access when they cannot verify the agent or its authority.
- Public evidence does not yet measure autonomous misconduct across the banking sector.
No reputable public source has confirmed a sentient machine choosing to raid bank accounts. The known problem involves identity, authority, software behavior, and attacker misuse.
Which Incidents Get Confused with Rogue Agent Attacks?
Several well-sourced incidents show that agents can perform dangerous cyber actions. They do not all represent the same threat. Combining them produces a more dramatic story, but a less accurate one.
| Evidence | What happened | Who was affected | Correct classification |
|---|---|---|---|
| Bank account access reports | Third-party agents entered customer sessions or attempted account actions | Direct | Identity and authorization failure under investigation |
| UK AISI evaluation | Agents took unsanctioned actions on the live internet | Real people and an open-source project; no harm found | Controlled test with real external effects |
| Anthropic threat report | An attacker used agents across a cyber espionage campaign | Tech companies, financial firms, manufacturers and government agencies | Attacker-directed agentic cyberattack |
| IMF and BIS analysis | Researchers modeled faster attacks and wider spillovers | Financial sector | Systemic risk analysis, not an incident report |
The UK Evaluation Tested Autonomous Behavior
The UK AI Security Institute ran 122 cyber evaluation attempts. Agents took 19 unsanctioned actions across ten runs. One agent attempted a malicious open-source contribution and social engineering.
The test used internet access and disabled provider safeguards. Those conditions do not match normal public deployments. AISI found no resulting real-world harm and reported no attack on banking systems.
The test still offers a practical warning. An agent may pursue a goal through routes its operator never approved. Internet access and weak containment can turn unexpected behavior into external action.
The Cyber Espionage Campaign Had a Human Attacker
Anthropic reported a state-sponsored campaign targeting about 30 organizations, which succeeded in a small number of cases. Targets included large tech companies, financial institutions, chemical manufacturers and government agencies. The attacker used Claude Code for 80% to 90% of tactical work.
Human operators selected targets and approved several decision points. They also deceived the model about the work's purpose. This was an AI-orchestrated attack, not an independent model choosing targets.
The campaign still changes threat planning for every organization. Agents can scan systems, find weaknesses, write exploit code, and collect credentials quickly. Human attackers can direct that work with fewer interventions.
What Makes Agent Access Different from Ordinary Automation?
Traditional software follows coded paths. An agent can interpret goals, select tools, read external content, and plan several steps. That flexibility also creates more ways to cross an authorization boundary.
A browser agent may use the user's password. The service then sees the user's authenticated session. It may not see a distinct software identity or a clear delegated mandate.
NIST warns that shared and long-lived credentials weaken attribution. Whoever holds a bearer token may present it. Broad API keys also provide weak proof of the acting party's identity.
The main risk paths include:
- Credential sharing. The agent receives the user's full account access.
- Prompt injection. Malicious content changes the agent's instructions.
- Tool misuse. The agent uses an allowed tool for an unapproved action.
- Privilege escalation. A low-trust task reaches higher-value systems or data.
- Memory poisoning. Harmful instructions persist into later sessions.
- Approval manipulation. The interface hides context from the human reviewer.
OWASP's AI Agent Security guidance also flags irreversible actions, such as payments, deletions and deployments. It recommends separating decisions from execution for such actions.
Could an Agent Attack a System Without a Human Attacker?
The capability exists in limited and specific settings. Current public evidence does not establish routine autonomous attacks on companies. The strongest examples involve permissive testing, compromised instructions, broad tool access, or human-directed campaigns.
That answer can change as models and tools improve. Security teams should plan from demonstrated capabilities, not science-fiction labels. The safer assumption is that agents increase speed and reduce attacker effort.
The Bank for International Settlements identifies the compressed remediation window as a core risk. Frontier models can find vulnerabilities and chain exploits faster. Defenders receive less time to patch exposed systems.
The IMF's 2026 financial-sector note focuses on scale. Shared cloud services and software can spread an incident across many firms. The concern is speed and reach, not a new magical attack method.
Three scenarios deserve separate planning:
- An attacker directs an agent against a company or service provider.
- Malicious content hijacks an authorized agent, such as a coding or email assistant.
- An agent exceeds its mandate without an external attacker.
Each scenario needs separate detection rules and evidence. A single "rogue" label hides those differences.
Why a Valid Login Is Not Valid Consent
Authentication proves that a credential worked. It does not prove that every later action matches the user's instruction. Agentic access turns that old distinction into an operational problem.
A customer may authorize an agent to read balances. The same session might expose payment, password, and payee controls. A stolen instruction could then trigger a much broader action.
The Federal Reserve recommends explicit consent for specific actions. Barr also called for strict authentication and technical limits on agent capabilities. Those controls preserve more context than a shared password.
Existing interagency guidance already supports layered controls. It covers customer-permissioned access, logging, anomaly detection, and stronger authentication for risky transactions.
For any service an agent acts in, useful evidence should answer five questions:
- Which agent initiated the request?
- Which person or organization controls that agent?
- What exact action did the user approve?
- Which tools and data could the agent reach?
- Can the service revoke access immediately?
Without those answers, incident teams may see a valid session with invalid intent.
How Does Risk Change as an Agent Gains Financial Authority?
Agent risk rises at each step from reading data to changing who has access. Each level needs different controls. A single "agent access" permission hides the difference between reading a repository and changing its admin settings.
The following model is Stack Row's synthesis of OWASP, NIST, FINOS, HKMA and Federal Reserve guidance. Each row gives a software and a banking example.
A Five-Level Agent Authority Ladder
The ladder classifies access by the most consequential action available. Higher levels require stronger identity, approval, monitoring, and recovery controls.
| Level | Agent capability | Software example | Banking example | Minimum control |
|---|---|---|---|---|
| 1. Observe | Read data | Read a repository or log | Check a balance | Scoped, read-only token |
| 2. Recommend | Suggest an action | Propose a fix | Compare savings rates | Clear sources and disclosures |
| 3. Prepare | Draft a pending change | Open a pull request | Draft a bill payment | Human reviews every final parameter |
| 4. Execute | Commit the change | Merge and deploy | Send a payment | Action-bound approval and limits |
| 5. Administer | Change future authority | Grant admin rights or rotate keys | Add a payee or reset access | Separate channel and stronger verification |
Level five can create more risk than any single action. An access change may enable many later actions, so score authority changes separately from the size of any one change or payment.
The Mandate Envelope
A valid mandate should describe more than the user's identity. It should define the complete boundary around each authorized action.
That boundary should record:
- The verified customer and identified agent.
- The permitted account, product, and communication channel.
- The approved action and blocked actions.
- The amount, recipient, frequency, and cumulative limit.
- The start time, expiration, and revocation method.
- The required confirmation for material changes.
- The evidence retained for disputes and investigations.
The Hong Kong Monetary Authority's discussion paper supports this approach. It recommends defined mandates, transaction limits, clear records, and prompt revocation.
This model reveals a common design error. Teams authenticate the customer once, then let the model interpret later authority. Deterministic policy should validate every consequential tool call.
Which Controls Reduce Rogue Agent Risk?
Organizations need agent-specific controls layered onto established cyber defenses. Model guardrails alone cannot enforce every downstream permission. The surrounding identity, tool, and transaction systems must set the boundary.
The FINOS AI Governance Framework identifies agent action authorization bypass as financial risk RI-24. It covers tool chaining, API discovery, and approval-workflow circumvention.
Give Every Agent a Separate Identity
Agents should not inherit a person's full username and password. They need scoped, short-lived, and revocable credentials. Permissions should identify the agent, user, purpose, payee, amount, and expiration.
Read-only access should remain separate from transaction authority. A customer should approve each high-risk expansion. Services also need a reliable way to reject unidentified browser automation.
Bind Approval to the Exact Action
Generic confirmation screens provide weak protection. Approval should cover the final amount, destination, timing, and account. Any material change should invalidate the approval.
High-risk changes need a separate verification path. Examples include new payees, password resets, contact changes, and large transfers. The OCC recommends out-of-band confirmation and velocity monitoring for emerging fraud risks.
Limit Tools and Contain Failures
An agent should receive only the tools required for its current task. Payment tools should remain unavailable during a balance inquiry. Administrative tools should never share the same trust boundary.
Useful containment controls include:
- Network allowlists for approved endpoints.
- Transaction limits tied to the agent's mandate.
- Sandboxed execution for code and browser tools.
- Circuit breakers for unusual tool chains or request volume.
- Immutable logs for prompts, decisions, calls, and approvals.
- Automatic revocation after anomalies or policy violations.
These controls also support faster investigation. Stack Row's cybercrime statistics guide explains why overlapping incident labels can distort reporting. Clear event categories help teams measure the real exposure.
Test the Deployed System, Not Only the Model
Agent behavior depends on tools, instructions, memory, and connected data. A model vendor cannot reproduce every deployment. Each organization needs tests that reflect its own permissions and workflows.
OWASP recommends repeated tests for prompt overrides, tool misuse, and approval bypass. Teams should also test memory poisoning and multi-agent chaining. Material configuration changes should trigger another test cycle.
Teams can compare these findings with wider cybersecurity statistics by sector. The comparison helps separate agent-specific failures from familiar access and patching problems.
Who Owns Each Control When Several Parties Share the Transaction?
No single party controls the entire agent action. The service, such as a SaaS platform, cloud provider or bank, governs execution. Agent providers govern agent security. Users define mandates, and protocol operators protect the communication channel.
This responsibility map adapts the HKMA's proposed shared-responsibility model.
| Party | Primary responsibility | Evidence it should retain |
|---|---|---|
| Service (SaaS platform, cloud provider or bank) | Validate authority and enforce action boundaries | Agent identity, mandate, checks, decision, and result |
| Agent provider | Protect credentials and report compromises | Agent version, tool policy, security events, and notices |
| User or customer | Grant, narrow, and revoke authority | Approved purpose, limits, duration, and confirmations |
| Protocol or intermediary | Protect message integrity and channel records | Signed requests, delivery events, and protocol changes |
Shared responsibility cannot become diluted responsibility. The service still controls whether its systems accept and execute an instruction. The provider remains responsible for the agent's security and truthful identity claims.
This split also improves incident response. Investigators can request the correct evidence from each party. Customers receive a clearer path for disputes and access revocation.
What Should Users Do Before Giving an Agent Account Access?
Avoid sharing passwords with general-purpose agents, whether for email, cloud consoles or bank accounts. Credential sharing removes many identity and revocation protections. It may also affect responsibility for resulting transactions.
Use these checks before connecting any financial account:
- Confirm that the service supports the connection.
- Prefer the service's own authorization page (OAuth) over password sharing.
- Choose read-only access unless the task truly needs more.
- Review the exact scopes, accounts, limits and duration.
- Enable alerts and strong multifactor authentication.
- Revoke access when the task ends.
- Report any unexplained agent activity immediately.
Do not assume a familiar agent name proves the service approved it. Check the service's current policy and connection method.
What the Research Says Now
The rogue agent story is serious, but narrower than many headlines suggest. Agents are reaching accounts, repositories and cloud consoles faster than services can standardize identity and consent. Separate cases show that agents can perform long cyber sequences at machine speed.
The missing fact remains important. Public sources have not confirmed widespread self-directed agent attacks on major companies. They have confirmed the conditions that could make such attacks harder to attribute and contain.
Stack Row's AI statistics track how fast companies are putting AI into production, and the explainer on what AI agents are covers tools, permissions and human approval in more depth. The technology question hub tracks practical answers as the evidence changes. Security teams should review agent access whenever tools, permissions, or payment functions change.
Frequently Asked Questions
What Is a Rogue AI Agent?
A rogue AI agent acts outside its approved goal, authority, or effective oversight. Causes include manipulation, unsafe permissions, software failure, or unexpected goal pursuit.
Have AI Agents Hacked Companies on Their Own?
Not in any confirmed public case. Agents took unsanctioned actions in a government test, attackers used agents in a cyber espionage campaign, and banks have blocked unverified agents in customer accounts. None of these is an agent independently breaching a major company.
Was the Optimizely Attack on Zoom, BBC and IKEA Websites Done by AI?
No evidence shows that. Scott Helme, who found the attack, said researchers saw no evidence of AI involvement. It was a supply-chain attack: one trusted vendor's scripts were tampered with, so the code ran on many customers' sites at once.
Was the November 2026 Apple Outage a Cyberattack?
No. Apple said an internal update caused the November 17 outage, which hit Apple Account sign-in, iCloud, the App Store, Apple Music and Apple Pay. No report linked it to an attack or to AI agents.
Were the Recent Bank Outages Caused by AI Agents?
No. Chase said the November 16 problems that also affected other banks and Zelle came from a technical bug, not a cyber issue. The October Google Cloud outage that disrupted UK banks came from a power system misconfiguration.
Why Are Banks Blocking Third-Party AI Agents?
Banks may not identify the agent or verify its precise mandate. Shared credentials also weaken attribution and revocation.
Can Prompt Injection Make an Agent Take Harmful Actions?
Yes. Prompt injection can redirect an agent that reads hostile content, such as a web page, email or shared file. Deploys, deletions, payments and permission changes should still require separate, action-bound approval and enforced limits.
Is Read-Only Access Safe for an AI Agent?
Read-only access lowers the risk of changes but still exposes sensitive data, such as source code, email or financial records. Use scoped authorization and revoke access after the task.
Which Standard Covers AI Agent Security?
No single standard covers every use case. NIST, OWASP, financial regulators, and sector guidance provide complementary controls.
Resources
- The Register: Google's October 20 London outage and its generator misconfiguration
- CIO.com: Microsoft struggles to get M365 services back online
- CBS News: Widespread service outages at major US banks
- 9to5Mac: iCloud, App Store and other Apple services hit by outage
- Engadget: iCloud and other Apple services are down
- Federal Reserve: Potential Financial Stability Implications of Artificial Intelligence
- Reuters: US Banks Blocking Third-Party AI Agent Access
- UK AISI: Unsanctioned Agent Behavior During Cyber Testing
- Anthropic: AI-Orchestrated Cyber Espionage Campaign
- NIST: Security Considerations for AI Agents
- OWASP: AI Agent Security Cheat Sheet
- IMF: Artificial Intelligence and Cybersecurity in the Financial Sector
- BIS: When Machines Attack
- OCC: Semiannual Risk Perspective, Fall 2026
- HKMA: Discussion Paper on Agentic AI Governance in Banking
- FINOS: AI Governance Framework
