The OCC's examiner framework (Bulletin 2026-54) says an AI failure includes a system that keeps running while acting wrongly, so a bank may plan how to detect, contain, recover from and fix it while critical work continues. It names kill switches, circuit breakers, rate limits and least-privilege permissions as containment tools. On security it names five AI-specific risks, from prompt injection to agents with too much access.
TRACKED DOCUMENTS: OCC Bulletin 2026-54
What does operational resilience mean for an AI system?
It means the bank can keep its critical work going when an AI system fails, and can spot, stop and repair the failure. The OCC framework adds one idea that ordinary IT resilience does not stress: an AI system can fail while it is still running.
The OCC calls AI “Super Intelligence (SI)” in this framework. This page writes AI in its own sentences and keeps SI only inside quotations.
The framework groups its sixteen parts into six pillars (section 1.2). The Resilience pillar “addresses the bank's ability to detect, contain, recover from, and remediate SI system or agent failures while continuing critical business processes.” Two components sit under it: Operational Resilience (section 2.10) and Cybersecurity (section 2.11).
Section 2.10 opens with this: “Operational resilience for SI includes both loss of service and incorrect operation.” Loss of service is the familiar case. The system is down. Incorrect operation is the new case. The framework says “an agentic system could remain available while taking inappropriate or harmful actions”. An agentic system is AI that can take actions on its own, not only write text.
For example, a chatbot that crashes is annoying but visible. An agent that keeps running and quietly waives the wrong fees for a week is harder to notice, and the harm builds up. The framework's summary table names this exact risk for section 2.10: “A system that remains available while acting incorrectly, with no tested containment, continuity, recovery, or remediation path”.
Like the rest of the framework, this is a resource for OCC examiners. Bulletin 2026-54 says it “does not establish supervisory expectations for banks”. It describes what examiners may look at, not rules a bank has to follow.
What does the OCC framework say about detecting, containing and recovering from an AI failure?
Section 2.10 lists the steps a bank may document and test for its more serious AI uses: “detection, containment, suspension, continuity, recovery, escalation, and remediation arrangements.” It then gives examples for each stage.
The framework says this applies “for more consequential or autonomous use cases”. Put simply, a low-risk drafting tool is not the focus. Which uses count as consequential is the subject of the OCC AI risk tiers page.
| Stage | What section 2.10 lists | Put simply |
|---|---|---|
| Detect and contain | “Monitoring, observability tools, control agents, deterministic rules, circuit breakers, rate limits, least-privilege permissions, and kill switches” could help identify and contain an agent “that is operating but acting incorrectly.” | Notice the agent is misbehaving, then limit or stop it. |
| Keep the work going | The bank may test whether the process can continue “through trained personnel, a narrower fallback agent, or another approved process” while the main system is restricted or unavailable. | Someone or something else does the job while the AI is off. |
| Recover and review | “Rollback, data correction, restoration of decision support, root-cause analysis, model-risk review, and controlled return to service” could be part of recovery exercises. | Undo the damage, find out why, and switch the AI back on carefully. |
| Fix it for customers | When a misfire affects customers, a separate workstream “to identify the complete affected population, assess individual harm, provide notice or correction, and independently verify completion.” | Find every customer who was hurt, put it right, and have someone else check it was done. |
A few of these terms need a plain explanation. A circuit breaker is an automatic stop that trips when activity crosses a set limit. A rate limit caps how many actions the system can take in a given time. Least-privilege permissions means the agent gets only the access it needs for its job and nothing more. A control agent is a second automated system that watches the first one. Deterministic rules are fixed rules that always give the same answer for the same input. A misfire is a word the framework uses in sections 2.3, 2.9 and 2.10 without defining it. Put simply, it is an AI action that goes wrong.
In practice, picture an agent that releases fraud holds on cards. A rate limit could stop it releasing more than a set number an hour. A circuit breaker could pause it if releases jump far above normal. The fraud team could then work the queue by hand while the cause is found. That example is ours, not the OCC's.
What is a kill switch, and what does the OCC say about it?
Put simply, a kill switch is a way to shut an AI system or agent off quickly, so it stops taking actions. The framework names kill switches once, in section 2.10, as one of several tools that “could help identify and contain an agent that is operating but acting incorrectly.”
That is all the framework says about kill switches by name. It does not define the term, does not say every system needs one, and does not say how fast one should work. The idea of stopping an agent appears in other places in different words:
- Section 2.1 says a bank may document “the circumstances that may require human review or suspension” of a system before it goes live.
- Section 2.3 says the assigned reviewer may need the expertise and authority “to evaluate, reject, modify, pause, or halt the system's output or action.”
- Section 2.4 lists suspension among the decisions for which roles and decision rights may be defined.
- Section 2.12 says a bank may define thresholds that trigger “review, re-validation, reduced autonomy, or suspension.”
A kill switch only helps if the work can carry on without the AI. That is why section 2.10 pairs containment with business continuity. The bank may test whether the process continues through trained staff, “a narrower fallback agent”, or another approved process. A narrower fallback agent is a simpler AI system with fewer permissions that takes over a smaller part of the job.
For example, if a bank switches off an AI agent that answers payment disputes, staff who still know how to handle disputes by hand can keep the queue moving. If nobody has done the task manually for a year, the switch is harder to use. Section 2.15 adds an assurance angle: when a fallback agent, circuit breaker or human process serves as a backstop, internal audit may assess whether it is “independently governed, tested, and effective.” Who decides to pull the switch is covered on the human oversight page.
Does the OCC framework raise dependence on one AI provider or model?
Yes, but as concentration risk spread across several sections, not as a separate resilience topic. Section 2.10 itself does not mention a single provider.
- Section 2.2 says “shared providers, data, permissions, infrastructure, and control weaknesses could create concentration or correlated risk across otherwise separate use cases.”
- Section 2.9 says reliance on third parties can bring “concentration, and resilience risks”. Its summary-table risk includes “switching constraints, or concentration in common models and infrastructure.”
- Section 2.10, under stress and systemic effects, says scenario analysis may consider “correlated or cascading activity across customers, agents, and providers.”
- Section 2.16 lists “reliance on significant providers” among the things reporting may help supervisors understand.
Put simply, if five of a bank's AI tools run on the same outside model, one outage or one bad update at that provider can hit all five at once. Switching constraints means it may be hard to move to another provider quickly. Contracts, exit plans and provider agents are covered on the vendor and third-party AI page.
What about AI agents the bank does not control?
Section 2.10 treats outside agents that reach the bank as a channel risk. It lists controls that “may be calibrated to the action requested”.
The text: “Customer, counterparty, fintech, or agentic-commerce systems that the bank does not control could present channel risk. Identity, provenance, delegation, authentication, monitoring, velocity limits, dispute handling, suspension, and revocation may be calibrated to the action requested.” Delegation means checking that a person really gave the agent authority to act for them. Velocity limits cap how fast or how often an agent can do something. Revocation means withdrawing that agent's access.
For example, a customer's shopping agent asking for a balance is a small request. The same agent asking to move money to a new payee is a bigger one, and could face stronger checks. Section 2.10 also warns that “high-speed agent activity could affect liquidity, fraud losses, service availability, or market behavior.” How large banks currently admit or block outside agents is tracked on our agent access page.
Which AI and agent security risks does the OCC framework name?
Section 2.11 names five. They share one point: AI systems can be attacked through their inputs, data and permissions, in ways that firewalls and ordinary code reviews were not built to catch.
The section says AI systems “may be targets of cyberattacks and tools that increase the speed, scale, or sophistication of attacks.” It adds that “prompts, model inputs, training and fine-tuning pipelines, connected tools, and non-human identities create attack surfaces that may not be fully addressed by controls designed for conventional code and network perimeters.” An attack surface is every point where an attacker could get in. A prompt is the text instruction given to an AI system.
- Prompt injection and jailbreaking. The framework: “Malicious or misleading inputs can alter system behavior, bypass restrictions, or cause an agent to misuse connected tools without a conventional network breach.” Prompt injection hides instructions in text the AI reads. Jailbreaking talks the AI out of its own safety rules. For example, a loan application PDF could contain hidden text telling a review agent to mark the file as approved.
- Data and model poisoning. Compromised training, fine-tuning, verified source or feedback data “could embed undesirable behavior that is difficult to detect after deployment.” Put simply, someone corrupts what the AI learns from. For example, an attacker who feeds false “not fraud” labels into feedback data could teach a fraud model to let a pattern through.
- Model extraction and data leakage. “Repeated queries or crafted inputs could be used to infer model behavior, recover sensitive training information, or expose confidential data.” For example, someone could ask a customer chatbot thousands of careful questions to learn its credit rules, or coax it into repeating another customer's details.
- Tool use and excessive agency. “Broad credentials or permissions could allow a compromised or manipulated agent to cause fraud, data exposure, or operational disruption at machine speed.” Excessive agency means an agent can do more than its job needs. For example, an agent built to read account balances that also holds rights to send wires is a bigger prize for an attacker.
- Non-human identities and permissible use. Agent identities, credentials, logs and data use “may be governed in a manner comparable to privileged users”. A non-human identity is a login used by software rather than a person. In practice, the agent's account would get the same scrutiny as a system administrator's.
The examples in this list are ours. The framework names the risks but gives no scenarios for them. The lexicon in Appendix A also defines adversarial SI as techniques and attacks that manipulate AI systems into wrong or unintended decisions, and data poisoning as an attack that corrupts training data. Testing for these risks (adversarial testing and red teaming, section 2.13) is covered on the model risk management page.
What does the OCC framework say about access rights for AI agents?
It treats an agent's access as part of its risk. Section 2.11 says “the breadth of agent permissions and tool access may be treated as part of the attack surface”, and section 2.10 lists least-privilege permissions as a containment tool.
The same idea appears across the framework:
- Section 2.1: before deployment, a bank may document “the data and systems it is permitted to access” and “the actions it could take”, stated “in terms that can be tested”.
- Section 2.2: “the breadth of its data, tool, and account access may materially affect risk.”
- Section 2.6: technical controls may limit the system to approved data sources and prevent access outside its purpose.
- Section 2.9: for provider agents, a bank may define “the scope of authority, identity, permitted tool use, data use and escalation limits”.
- Section 2.14: permission changes may be logged and linked to testing and approval evidence.
In practice, this works like access for a new employee. A teller does not get the keys to the vault. An agent that drafts replies to customers would not get rights to change account records. The agent lifecycle page shows where these checks fit when a bank builds an agent.
Which control does the framework describe for each named risk?
Section 2.11 names risks but never pairs a risk with a control. The pairing in this table is ours. Each control comes from the section cited in its row.
| Risk named in the framework | What it means | Related controls in the framework (our pairing) |
|---|---|---|
| Incorrect operation while running (2.10) | The AI stays up but takes wrong actions. | Monitoring, control agents, circuit breakers, rate limits, least-privilege permissions, kill switches (2.10) |
| Loss of service (2.10) | The AI is unavailable. | Fail-over to trained personnel, a narrower fallback agent or another approved process (2.10) |
| Outside agents (2.10) | Agents the bank does not control act through its channels. | Identity, authentication, delegation, velocity limits, suspension, revocation (2.10) |
| Prompt injection and jailbreaking (2.11) | Hidden or tricky inputs hijack the AI. | Adversarial testing of prompt, data and tool controls, and red teaming (2.13) |
| Data and model poisoning (2.11) | Corrupted learning data plants bad behavior. | Data and model integrity controls (2.11), data lineage review (2.6) |
| Model extraction and data leakage (2.11) | Repeated questions reveal how the model works or what it saw. | Monitoring and identity and access management (2.11), access controls on prompts, outputs and logs (2.7) |
| Tool use and excessive agency (2.11) | An agent with too much access does damage fast. | Treating permissions as attack surface (2.11), least privilege (2.10), testable permitted actions (2.1) |
| Non-human identities (2.11) | Agent logins and credentials are a target. | Govern them like privileged users (2.11) |
The key concepts of section 2.11 list the areas cybersecurity controls may extend to: “identity and access management, data and model integrity, secure development, monitoring, incident response, and recovery.” Each component in one paragraph is on the sixteen components page.
How does this connect to AI incident response?
Resilience and cybersecurity are the parts of the framework that cover what happens during and after an AI incident. Other sections cover how the incident is named, reported and checked.
- Section 2.11 lists incident response and recovery among the areas cybersecurity controls may extend to.
- Section 2.3 says a misfire, hallucination or control breach “may justify temporarily increasing human involvement.” A hallucination is a confident answer that is wrong.
- Section 2.9 says a bank may define “a specific SI incident vocabulary and notification timelines” to capture events such as hallucination, misfire, drift or manipulation.
- Section 2.14 lists incidents among the things documentation may cover, and decision traceability lets a bank reconstruct what an agent did.
- Section 2.16 says a bank may set internal escalation and notification thresholds based on “severity, duration, customer impact, financial exposure, legal requirements, and potential systemic effects.”
The framework sets no reporting deadline of its own. Section 2.16 notes that notifications required by laws, regulations and contracts “could be applicable to SI use cases.” Real cases of AI going wrong at financial firms are collected in our AI incidents database. Section-by-section depth is in the full Bulletin 2026-54 reference, and all OCC AI documents are on the OCC page.
Does the OCC require banks to have a kill switch for AI?
No. The framework is a resource for examiners and the bulletin says it does not set expectations for banks. Section 2.10 lists kill switches as one of several tools that could help contain an agent acting incorrectly.
Why is a running AI system a resilience problem?
Section 2.10 says resilience for AI covers both loss of service and incorrect operation. An agent can stay available while taking harmful actions, so being up is not the same as working.
What is a fallback for an AI system?
It is how the work continues while the AI is restricted or off. Section 2.10 names trained personnel, a narrower fallback agent, or another approved process.
Is prompt injection a cybersecurity issue for the OCC?
Yes. Section 2.11 lists prompt injection and jailbreaking first among its AI security examples. It says such inputs can make an agent misuse connected tools without a conventional network breach.
Should an AI agent have its own login?
The framework does not say so directly. Section 2.11 says agent identities, credentials, logs and data use may be governed like those of privileged users, and section 2.9 mentions identity for provider agents.
Does the framework cover customers' own AI agents?
Yes. Section 2.10 says customer, counterparty, fintech or agentic-commerce systems the bank does not control could present channel risk. Checks such as identity, authentication and velocity limits may be scaled to the action requested.
Does the framework set a deadline for reporting AI incidents?
No. Section 2.16 says a bank may define its own escalation and notification thresholds, and that notifications required by law, regulation or contract could apply to AI use cases.
- Super Intelligence: Supervisory Framework — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
- OCC Bulletin 2026-54: Super Intelligence: Supervisory Framework — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
When this rule moves, the next morning's brief says so
when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning