Bank of America Brought In 1,000 Contractors to Fix What AI Found. Scotiabank Named Mythos in Its Filings in May. Bank of America Still Hasn't.

2:10 AM Central

At 2:10 AM Central on Saturday, October 17, Cloudskope typed one word into the Securities and Exchange Commission's full-text search: Mythos. We limited the search to annual, quarterly and current reports filed since April 1, 2026, the week before Anthropic announced the model and the Treasury Secretary summoned the chief executives of America's largest banks to Washington to talk about it.

The search returned seventeen documents. Two were quarterly reports from a medical device company. Two were annual reports, from Ocean Power Technologies and NetApp. Most of the thirteen current reports came from cybersecurity vendors describing Mythos as a commercial tailwind; the others included Amazon, IBM and a biotech. Not one came from a bank, a broker, an insurer or any other filer in the finance, insurance or real estate codes.

At 2:32 AM we widened the search to every form type. It returned 34 filings. Four came from banks, and none of the four is American: two quarterly reports each from the Bank of Nova Scotia and the Canadian Imperial Bank of Commerce, furnished to the SEC on Form 6-K. Scotiabank told its investors in May that frontier AI such as Mythos was "materially increasing exposure across industry technology landscape." No U.S. bank appears anywhere in the results.

Three days earlier, on Wednesday, October 14, the chief executive of the second-largest bank in the United States had told analysts that his bank spent "a couple hundred million plus in the third quarter" on two remediation programs, and that one of them was "remediation for the Mythos implementation." Brian Moynihan said Bank of America had "added 1,000 extra staff to work on that," and then clarified where they sit: "not in the headcount, in consultants."

In a CNBC interview the same day, Moynihan said the contractors had come on in the second quarter, joining roughly 3,000 of the bank's own employees on the same work, and that the weaknesses being fixed, surfaced by Mythos and other frontier AI models, sat in code used by the bank and by its outside vendors. "That risk is real risk," he said.

Bank of America's second-quarter 10-Q, filed July 31, covers the exact quarter in which those 1,000 contractors arrived. Cloudskope read all 705,911 characters of its text. It contains the word cyber once, inside the boilerplate forward-looking-statements paragraph. It does not contain the words Mythos, vulnerability, contractor, consultant or remediation.

The 1,000 contractors are not the story. Bank of America can afford them, and by its own account the work is going well. The story is what the number tells everyone else. The most heavily resourced bank technology organization in the country, with a technology budget measured in billions of dollars a year, could not absorb what Mythos and similar models found in its code without renting a thousand more people. It did that quietly, in a quarter its securities filings describe without a word about it, in the same year in which its regulators were building the case that ended, on October 13, in more than $1.2 billion in penalties for knowing about a control gap for years and not fixing it fast enough.

Finding problems just became cheap. Fixing them did not. Every bank, credit union, wealth manager and portfolio company that does not have Bank of America's balance sheet is now on the wrong side of that gap, and most of their boards have not been told.

What was actually said, and where

Precision matters here, because most of the coverage this week compressed three separate statements into one headline.

The first statement is the earnings call transcript. On the call, Mike Mayo of Wells Fargo asked Moynihan about revenue per employee and, "in this period of AI and AI deployment," where that could take the bank. Moynihan's answer covered expenses: total expenses up about $1.2 billion from a year earlier, about $450 million of it wealth management incentives tied to revenue and about $300 million in markets. Inside that answer he said "a couple hundred million plus in the third quarter were two things": "remediation for the AML and remediation for the Mythos implementation." He said he expected the Mythos costs to "run down" because the 1,000 consultants were in place. Answering John McDonald of Truist a few minutes later about the OCC and FinCEN orders, he used the same order of magnitude for the anti-money-laundering effort: "a couple hundred million dollars plus in remediation costs for that effort."

Read the two passages together and the number does not split. Bank of America has not said how much of the couple hundred million went to Mythos and how much to anti-money laundering. Nobody outside the bank can say. Any reading that treats the full figure as the cost of the Mythos work goes beyond the transcript.

The second statement is the CNBC interview, where the timing (second quarter), the 3,000 existing staff and the vendor-code detail come from. Cloudskope could not retrieve the CNBC page directly; those details are as reported by CNBC and carried by PYMNTS and The Next Web, both of which attribute them to the interview. The vendor-code detail appears only in The Next Web's account, as a paraphrase.

The third statement is the earnings release itself, Exhibit 99.1 to the 8-K filed October 14. It reports noninterest expense of $19.7 billion, up $2.4 billion, and adjusted noninterest expense of $18.5 billion excluding the $1.2 billion regulatory settlement charge, against $17.3 billion a year earlier. It reports headcount of 209,911 at September 30, down from 211,304 at June 30 and 213,384 a year earlier. It does not mention Mythos, AI-identified vulnerabilities, remediation or contractors. The 1,000 consultants are, by Moynihan's own description, outside that headcount figure.

So the public record of Bank of America's Mythos program is two sentences on a call and one television interview. Everything else is inference.

What we verified

Cloudskope ran the following checks between 1:55 AM and 2:40 AM Central on October 17, using only public data and ordinary requests. The full log, with commands and UTC timestamps, is kept with this analysis.

• 2:09 AM, EDGAR filing index. Bank of America Corp. (CIK 70858) has filed no Item 1.05 (material cybersecurity incident) or cybersecurity-related Item 8.01 current report in 2026. The October 14 8-K carries Items 2.02, 7.01 and 9.01 only. Nothing in this analysis suggests a reportable incident occurred; the check establishes that the Mythos program has never been the subject of a filing.
• 2:10 AM, full text of four filings. We extracted the text of the 2025 10-K (filed February 25, 2026), the first-quarter 10-Q (May 1), the second-quarter 10-Q (July 31) and Exhibits 99.1 and 99.3 of the October 14 8-K, and counted terms. Mythos: zero in all of them. Vulnerabilit: zero in both 10-Qs, nine in the 10-K. Cyber: once in each 10-Q, in the forward-looking-statements paragraph. Contractor and remediation: zero in both 10-Qs.
• 2:10 AM, EDGAR full-text search across all filers. The phrase Mythos in 10-Q, 10-K and 8-K filings dated April 1 to October 17, 2026 returned 2 quarterly reports, 2 annual reports and 13 current reports (the annual-report hits were confirmed by an independent re-run of the search at about 2:40 AM). None was filed by a company with a standard industrial classification code beginning in 6 (finance, insurance and real estate). Filers included CrowdStrike, Zscaler, Fortinet, Rapid7, Tenable, Rubrik, Netskope, IBM and Amazon.
• 2:11 AM, Microsoft's own vulnerability data. We pulled every monthly security update document from Microsoft's public CVRF API for January 2025 through October 2026 and counted entries by assigning CNA. Microsoft-assigned CVEs: 180 in October 2025, 1,456 in October 2026. January through October: 1,107 in 2025, 4,418 in 2026, a ratio of 3.99. Across all CNAs in Microsoft's documents, the totals were 3,535 and 12,658. BleepingComputer reported 1,411 flaws for October; our count of Microsoft-assigned entries is 1,456, and we attribute the difference to counting method (revisions and non-product entries), not to an error on either side.
• 2:12 AM, CISA's Known Exploited Vulnerabilities catalog (version 2026.10.16). Additions from January 1 through October 16: 202 in 2025, 264 in 2026, an increase of about 31 percent.
• 2:13 AM, earnings release. Headcount and expense figures above confirmed in Exhibit 99.1.
• 2:32 AM, EDGAR full-text search, all form types. The same query with no form filter returned 34 filings. The only banks were the Bank of Nova Scotia (6-K, May 27 and August 25) and the Canadian Imperial Bank of Commerce (6-K, May 28 and August 27). Three other hits from a filer with a finance code were registration statements from Nakamoto Inc. in which Mythos is part of a shareholder's name, Mythos Orange1 Co. Ltd., not a reference to the model.

Two of those numbers belong side by side, because together they are the most honest summary of where the industry stands. Disclosed vulnerabilities in Microsoft's products alone roughly quadrupled. Vulnerabilities known to be exploited in the wild rose by about a third. The patch volume is the cost every bank is paying now. The exploitation curve is the risk that arrives when attackers have what defenders have. Anthropic's chief executive put a date on that in May: six to twelve months.

What we could not verify: how many vulnerabilities Mythos found at Bank of America, how many were critical, how many remain open, whether the 1,000 consultants had access to production systems or source code, which vendors' code was involved, and how the couple hundred million divides between Mythos and anti-money laundering. Bank of America can answer every one of those questions.

The bank told its shareholders this would happen

The most striking document in this story is not a new one. It is Bank of America's 2025 annual report, filed February 25, 2026, six weeks before Anthropic announced Mythos.

In its risk factors, the bank wrote that emerging technologies such as AI are expected to increase cybersecurity risks, and then said exactly how: "AI lowers the entry barriers to plan and execute cyberattacks, enables more personalized and harder to detect social engineering, and improves vulnerability discovery, which may result in the increased likelihood of exploitation and the speed, scope, scale, and sophistication of cyberattacks."

Improves vulnerability discovery. Somebody at Bank of America saw it coming and said so in a document signed by the chief executive and chief financial officer.

The same filing explains how the board is supposed to learn when that risk materializes. In the cybersecurity governance section incorporated into Item 1C, the bank says its board "receives quarterly reports on the performance metrics for the GIS Program and the performance of the Corporation's cybersecurity risk appetite metrics, including metrics on vulnerabilities and third-party cybersecurity risks and incidents, and is notified promptly if a Board-level cybersecurity risk limit is breached."

That sentence is the hinge of this analysis. Bank of America has a board-level cybersecurity risk limit. It reports vulnerability metrics against it every quarter. In the second quarter of 2026, the bank brought in 1,000 outside workers to fix vulnerabilities at the pace it judged necessary. Either the volume of AI-surfaced findings moved those metrics through a board-level limit, in which case the board was notified promptly and the program is a board matter, or it did not, in which case a remediation surge large enough to need a thousand contractors stayed inside management's appetite. Both answers are defensible. Shareholders have been given neither.

The 10-K also says something about vendors that reads differently now. "[W]e are at risk from critical third-party information security and open-source or proprietary software defects and vulnerabilities. We must rely on our third parties to detect and promptly report cybersecurity incidents, vulnerabilities and technology failures." According to The Next Web's account of the CNBC interview, the weaknesses being fixed were in code used by both the bank and its outside vendors. If Bank of America's tools found flaws in vendor code, then for those flaws the reporting ran the other way: the bank found them and told the vendor. Every other customer of that vendor is depending on the vendor to pass it on.

Two Canadian banks wrote it down

Bank of America was not alone in seeing the risk coming. What sets it apart is what happened after.

The Bank of Nova Scotia's second-quarter report to shareholders, furnished to the SEC as Exhibit 99.1 to a Form 6-K on May 27, says in its discussion of cyber risk: "Recent advancements in frontier AI (e.g., Anthropic's Claude Mythos Preview) enable rapid identification of complex and previously unknown vulnerabilities, materially increasing exposure across industry technology landscape." The same sentence appears in the third-quarter report furnished on August 25. Cloudskope read both.

The Canadian Imperial Bank of Commerce's quarterly reports, furnished on May 28 and August 27, say: "Advanced models, including Claude Mythos, could accelerate impacts across cyber, data and technology domains if risks are not properly addressed." The August version reads "Advanced models, including Claude Mythos and frontier models."

Neither Canadian bank said what it was spending. Neither said how many findings it had or whether it had hired outside help. They did the minimum: they put the name of the risk, and in Scotiabank's case the word materially, into the document investors read. Among the 34 EDGAR filings since April 1 that contain the word Mythos, theirs are the only four from banks.

Bank of America did more than either of them. It rented a thousand people. It wrote less.

The same week, the same failure

Read the Mythos disclosure next to what happened the day before it, and the week stops looking like two unrelated stories.

On Monday, October 13, the Office of the Comptroller of the Currency announced a $750 million civil money penalty against Bank of America, N.A., the largest in the agency's history. FinCEN announced a $1.21 billion penalty that credits the OCC amount, and OFAC announced a separate $11.6 million sanctions settlement. The bank admitted to willfully violating the Bank Secrecy Act. Regulators said more than $5 billion in suspicious activity went unreported or was reported late, and that more than $1 billion of it involved national security and other high-priority risks, including activity linked to Chinese money laundering networks and suspected trade-based money laundering tied to the U.S.-Mexico border.

The OCC's release is worth reading for its verbs. The bank operated a "deficient transaction monitoring system." It had a "systemic gap in its transaction monitoring program." It "knew for years that it was failing to escalate potentially suspicious activities." Its "BSA audit function failed to identify long-standing BSA concerns." And, more than a year after a December 2024 cease-and-desist order told it to fix all of this, it "still failed to escalate potentially suspicious activities to investigators."

Strip away the subject matter and that is a description of a detection system that produced findings faster than the institution could act on them, an escalation path that did not carry the findings to the people who could act, and an audit function that did not notice the backlog. It is, almost word for word, the risk every bank now carries in vulnerability management.

Mythos is a detection system. In the hands of a defender it produces findings, many of them real, some of them chained from minor weaknesses into serious ones, at a rate no human team was staffed for. Reuters reported in May that the few large banks with access were finding several hundred to thousands of low- to moderate-ranked vulnerabilities in bank technology and were patching in days rather than weeks. The detection side of the problem is solved for anyone with access. What remains is the side Bank of America just paid $1.2 billion for getting wrong in a different domain: escalation, staffing, and the distance between knowing and fixing.

In response to the penalties, Bank of America said it had invested billions of dollars in technology and assigned thousands of employees to combat financial crime, and that "our team is far along with and committed to completing the work required to address the issues raised by our regulators." Swap the subject and it would serve as a statement about the Mythos program. That is the point. The bank's regulators have now established, in the largest penalty in OCC history, what they think of an institution that knows about a gap and cannot close it at the pace the risk requires. It would be surprising if that standard stayed confined to anti-money laundering.

Less than a week earlier, American Express was fined $350 million over similar anti-money-laundering failures. The enforcement climate for remediation failures in large financial institutions is not theoretical this month.

April: the meeting at Treasury

The Bank of America disclosure is the latest entry in a six-month record that the coverage this week mostly left out.

Anthropic announced Claude Mythos Preview and Project Glasswing on April 7, 2026. The company said the model had "already found thousands of high-severity vulnerabilities," including some in every major operating system and web browser, and limited access to a group of launch partners: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia and Palo Alto Networks, plus more than 40 additional organizations. It committed up to $100 million in usage credits and $4 million to open-source security groups. "Given the rate of AI progress," the company wrote, "it will not be long before such capabilities proliferate."

That week, according to Bloomberg News, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell summoned the chief executives of Citigroup, Morgan Stanley, Bank of America, Wells Fargo and Goldman Sachs to Treasury's headquarters. Jamie Dimon of JPMorgan could not attend. The purpose, per people familiar with the meeting, was to make sure the banks understood the risks from Mythos and similar models and were taking precautions. Powell's presence signaled a systemic-risk concern. Every bank summoned is classified as systemically important. JPMorganChase was the only bank Bloomberg named as a Glasswing participant.

Brian Moynihan was among the chief executives summoned. Bank of America's first-quarter 10-Q was filed on May 1, about three weeks later. Cloudskope found no reference to Mythos, AI-discovered vulnerabilities or the Treasury meeting in it. That is not a violation of anything. It is a data point about what a systemically important bank's chief executive considered worth telling shareholders in writing after being summoned to Treasury over a cyber threat.

The same month, the Independent Community Bankers of America said it had engaged Treasury and White House officials and members of Project Glasswing to "elevate community bank concerns," and launched an AI task force of nearly 100 bankers from 39 states. ICBA did not say, and has not said publicly since, that its members gained access to the model. The Bank of Canada met Canadian lenders on the same subject. The Bank of England planned to meet bank and insurance executives.

May: the regulators paused the exams

On May 12, Reuters reported that the handful of large banks with Mythos access were sharing findings with smaller banks that lacked it, that the model could combine several lower-risk weaknesses into one high-risk vulnerability, and that banks were upgrading end-of-life technology. Dario Amodei said on May 5 that financial services and other organizations had six to twelve months to fix software vulnerabilities before models from China matched Mythos, and that "we could be in a better position than we started in because we fixed all these bugs." Bessent said on May 3: "What we've had in the past month was a step change in the power of one large language model."

Anthropic said in May that Mythos Preview had found more than 10,000 vulnerabilities across the roughly 50 partner organizations with access at the time.

On May 19, Bloomberg News reported that the Federal Reserve and the OCC had paused some cyber-related examinations of the largest banks, to give them time to strengthen systems against threats exposed by the latest AI models. The OCC declined to comment. Neither agency has published the scope, the duration or the conditions of that pause.

Hold that next to Bank of America's timeline. The contractors arrived in the second quarter, which ended June 30. The exam pause was reported in the middle of it. The bank was, at the same time, operating under a December 2024 cease-and-desist order on anti-money laundering that its examiners would later find it had not satisfied. Bloomberg's report did not name the banks covered. If Bank of America was among them, as its size suggests, then a bank under an open enforcement order received, in another domain, a supervisory grace period to fix what an AI model had found. Whether the two supervisory teams compared notes is a question for the OCC. It is not a question anyone has asked in public.

June: an executive order that names community banks

On June 2, President Trump signed Executive Order 14409, Promoting Advanced Artificial Intelligence Innovation and Security. It directed Treasury, with NSA and DHS, to stand up an AI cybersecurity clearinghouse within 30 days to coordinate vulnerability scanning, validate findings and prioritize remediation and patch distribution. It named "rural hospitals, community banks, and local utilities" as intended beneficiaries of expanded access to cybersecurity tools and services. It allowed developers of qualifying frontier models to give the government up to 30 days of pre-release access, and it said it created no licensing or preclearance regime.

On July 14, the White House announced the clearinghouse, named GOLD EAGLE, run with Treasury, CISA and the Department of War, and said it had "begun to intake and prioritize identified cybersecurity vulnerabilities from across industries and sectors." Bessent said Treasury was working with the private sector to "safeguard our financial institutions, close vulnerabilities, and protect the integrity of the U.S. financial system." Participation is voluntary. As of a July 29 analysis by K&L Gates, the administration had not published guidance on how companies engage, what confidentiality protections apply to submitted vulnerability data, or how prioritization works. Cloudskope found no public guidance since that explains how a community bank, the institution the order names, actually gets in the door. CISA did issue Binding Operational Directive 26-04 on June 10, which requires federal agencies to remediate the highest-risk tier of vulnerabilities, publicly exposed, actively exploited, automatable and high-impact, within three days.

Three days is now the federal government's standard for its own systems. It is a useful number for a bank board to hold up against its own patching policy, and an uncomfortable one for most of them.

October: tenfold

On October 6, Jamie Dimon told Bloomberg Television that AI-related threats "went up 10-fold after Mythos," and that "AI created vulnerabilities that we didn't know about, and we always worried about cyber before these things." JPMorgan, he said, was "rolling up our sleeves and going to work to fix it." In May, according to The Next Web, Dimon had said JPMorgan had hundreds of people working on the issue full time.

On October 13, the day of the Bank of America penalties, Reuters reported that the Fed, the OCC and the FDIC were questioning banks about advanced AI: whether they use Chinese-developed technology, whether they use AI agents in cyber defense, whether they can monitor, approve, audit and shut down increasingly autonomous systems. The European Central Bank gave euro-zone banks until the end of October to produce plans against AI-enabled cyber threats, including mapping third-party dependencies. The Bank of England is encouraging banks to examine "bare metal" recovery plans that do not depend on their main technology stacks. European banks, Reuters reported, have complained about limited access to Mythos.

Then, on October 14, Bank of America put a headcount on the problem. It is among the first public accountings of what the work costs, and it came on an earnings call rather than in a filing.

This failure is nine years old, and the fix was 48 hours

None of this is new except the speed. The defining American financial-data breach of the last decade was a remediation failure, not a detection failure, and the public record on it is unusually complete.

On March 7, 2017, the Apache Software Foundation disclosed CVE-2017-5638, a critical remote code execution flaw in the Struts web framework, and published fixed versions. US-CERT notified Equifax the next day. On March 9, according to the House Oversight Committee majority staff report published in December 2018, Equifax circulated an instruction to patch affected systems within 48 hours. A scan on March 15 reported no vulnerable internet-facing systems; it did not reach the subdirectory where the Struts component in Equifax's online dispute portal lived. The portal stayed unpatched. Attackers entered it on May 13, 2017 (MITRE ATT&CK T1190, Exploit Public-Facing Application) and operated for about 76 days before an expired certificate on a monitoring device was replaced on July 29 and the traffic became visible. The House report puts the final count at about 148 million people.

Equifax was warned the day after the fix was published. It had a policy that said 48 hours. It had a scanner. The breach happened in the space between the finding and the fix, in a system nobody had mapped well enough to know it was in scope. Equifax's former chief executive, Richard Smith, later told Congress, as the House report records, that the failure traced to an individual who did not ensure the patch instruction reached the people responsible, and to a scan that did not find the vulnerable component.

Now multiply the number of findings by four, which is roughly what Microsoft's own data shows happened this year, and ask how many organizations have mapped their estate better than Equifax had in 2017. The industry has known since at least the Equifax hearings that discovery is not the constraint. Mythos made that impossible to ignore, and Bank of America just put a price on it.

The thousand-person side door

There is a second-order risk in the Bank of America number that no coverage this week addressed, and it is the one Cloudskope spends most of its time on.

Bringing 1,000 contractors into vulnerability remediation means granting 1,000 outside people, onboarded in a single quarter, access to the most sensitive engineering information a bank has: where its weaknesses are, which systems are unpatched, and often the source code and deployment pipelines needed to fix them. That is the textbook definition of MITRE ATT&CK T1199, Trusted Relationship, and every one of those accounts is a candidate for T1078, Valid Accounts. A list of a bank's open vulnerabilities is among the most valuable documents an attacker can hold. Remediation surges create a lot of copies of it.

On October 13, federal authorities announced Operation Lonestar Roundup: searches of more than 40 Texas residences and the seizure of 187 laptops and related devices used as laptop farms for North Korean IT workers. CrowdStrike, which worked with the FBI, said the broader campaign targeted more than 70 U.S. companies, and the organizations in the disrupted operations included financial services firms. "Each laptop seized potentially represents one company that has hired a DPRK IT worker," said R. Joseph Rothrock, special agent in charge of the FBI's Dallas field office.

To be explicit, as we are again below: there is no evidence and no suggestion that any Bank of America contractor is connected to that scheme. The point is structural. The industry's answer to AI-scale vulnerability discovery is human-scale contract staffing, procured fast, through staffing firms and consultancies, at exactly the moment a hostile state is running a contractor-infiltration operation that federal authorities are still unwinding. The interagency guidance on third-party relationships that the OCC, Fed and FDIC issued in June 2023 expects banks to scale due diligence to the criticality of the access. A contractor who can see the vulnerability backlog is about as critical as access gets.

Bank of America almost certainly has the vetting machinery to do this well. Most institutions copying the move will not.

Materiality runs downhill

For Bank of America, the Mythos program is almost certainly not material in the securities-law sense. A couple hundred million dollars, split with anti-money laundering, against a quarter with $30.0 billion in revenue and $6.9 billion in net income is a rounding item. The bank's silence in its filings is consistent with that judgment, and the judgment is probably right.

Now move the same problem down the size curve.

A $2 billion community bank or a mid-sized credit union does not have 3,000 people on this work to start with, let alone the budget to rent 1,000 more. It almost certainly does not have access to Mythos. What it has is a core banking provider, a digital banking platform, a card processor, a handful of managed service providers and a long tail of software vendors, and it depends on every one of them to find and fix the flaws in code it does not own. That is exactly the code Moynihan reportedly said Bank of America found weaknesses in: software used by the bank and by its outside vendors.

Follow that through. When a bank with Mythos finds a flaw in a vendor's product, the vendor fixes it for the bank that found it first. Every other customer of that vendor gets the fix when the vendor ships it, if the vendor ships it, and if the institution applies it. A finding that costs Bank of America a ticket costs a community bank an unknown period of exposure to a flaw that a well-resourced bank, and soon a well-resourced attacker, already knows exists. Reuters reported in May that large banks were sharing findings with smaller ones. There is no public description of how, with whom, or under what confidentiality terms.

This is Cloudskope's materiality asymmetry in its purest form. The same backlog that is immaterial to the institution that found it can be existential to the institutions downstream of it. The vendor that ships the fix will describe the episode, if it describes it at all, as routine maintenance. The community bank's board will not hear about it unless someone asks.

For private equity, the asymmetry runs through every portfolio. A sponsor-backed software company that sells into banks is on the receiving end of AI-scale findings from its largest customers, whether it has budgeted for it or not. A sponsor-backed services business with no frontier-model access is on the far side of the same gap as a community bank. Neither will show up in a quality-of-earnings report.

What the numbers say about everyone else

Microsoft's monthly update documents are the cleanest public proxy for what the AI-discovery wave looks like from the receiving end, because Microsoft publishes them in machine-readable form and assigns most of its own CVEs.

From January through March 2026, Microsoft assigned an average of about 94 CVEs a month to its own products. April, the month Mythos was announced, came in at 181. Then 162 in May, 220 in June, 663 in July, 456 in August, 998 in September and 1,456 in October. The October 2025 figure was 180. Including the third-party components Microsoft tracks in its documents, such as Chromium and Linux packages, October 2026 carried 2,135 entries and September 2,779.

Every one of those is a patch someone has to test and deploy. A bank that patched against a monthly cadence of a few hundred entries is now looking at two to three thousand entries a month in a single vendor's update documents, before counting Oracle, which shipped 1,449 patches in July according to The Register, or anyone else.

The other side of the ledger is the reason not to panic. CISA's catalog of vulnerabilities known to be exploited in the wild grew by 264 entries from January 1 through October 16 this year, against 202 in the same period of 2025. Exploitation is rising, but nowhere near as fast as disclosure. Defenders, for now, are finding faster than attackers are using. That is what a head start looks like. It is also what Amodei's six-to-twelve-month warning was about, and the clock he described in May runs out between November and next spring.

The hypothesis, stated as one

What follows is Cloudskope's inference, labeled as such.

Bank examiners have spent two decades refining how they treat a detection system whose findings outrun the institution's capacity to act: they call it an alert backlog, they cite it as a deficiency, and, as of October 13, they fine it at record levels when it persists after an order. AI-assisted vulnerability discovery turns every bank's vulnerability management program into exactly that kind of system. The exam pause in May bought the largest banks time. It will not last indefinitely, and when it ends we expect examiners to ask the anti-money-laundering question about vulnerabilities: what did you know, when did you know it, how long did it sit, and who above the first line was told.

If that is right, the record that matters is not the patch count. It is the age of the open findings, the escalation path, and whether the board's risk limit was set before or after Mythos. Bank of America's 10-K says it has such a limit and reports against it quarterly. In Cloudskope's experience, many institutions do not have one written down.

What Bank of America should have done by now

1. Disclose the program in writing. A two-sentence answer on an earnings call is not a disclosure. A paragraph in the third-quarter 10-Q, due within weeks, describing the scale of AI-identified findings, the remediation approach and the cost, would cost the bank nothing it has not already said aloud and would give every smaller institution a benchmark.
2. Split the number. The couple hundred million combines a regulatory remediation the bank was ordered to perform with a voluntary security program it chose to run. Investors, regulators and peers would read them very differently. One sentence separates them.
3. Say whether the board-level risk limit was breached. The 10-K promises prompt board notification if it is. Shareholders are entitled to know whether a remediation surge of this size triggered the mechanism the bank describes as its safeguard.
4. Describe the contractor controls. Not the names. The access model: whether consultants work from bank-managed devices, whether they can see the full vulnerability inventory or only their assigned items, how identity was verified at onboarding, and how access is withdrawn as the surge winds down.
5. Publish the vendor pathway. If Bank of America's tools found flaws in vendor code, the bank holds information that other institutions using the same vendors need. A public statement of how findings are passed to vendors and on what timeline, without naming vendors, would set an industry norm.
6. Commit to the October call, in writing, for the rest of the industry. Moynihan said the Mythos costs will run down. If the bank is willing to report that trajectory quarterly, it becomes the first public time series of what AI-scale remediation costs. Nobody else has one.

What boards should be asking this week

These questions apply to bank and credit union boards, to the boards of companies that sell into financial services, and to any board whose company runs software it does not write.

How many open vulnerabilities do we have today, how old is the oldest critical one, and what was each number on April 1? Do we have a board-level risk limit on vulnerability age or volume, and was it set before or after the AI-discovery wave began? Has management told us, in writing, whether we have access to a frontier vulnerability-discovery model, and if not, which of our vendors do? For our core, digital banking and payments providers: have they received AI-identified findings from larger customers, how many, and what is their remediation commitment to us? If we brought in outside help to clear a backlog, who can see the full list of what is open, and how did we verify who they are? What is our patch window for an internet-facing, actively exploited flaw, and how does it compare with the three days the federal government now requires of itself? And the uncomfortable one: if our examiners asked the anti-money-laundering questions about our vulnerability backlog (what we knew, when, how long it sat, who was told), would we like our answers?

Ask for the answers in a week. The time it takes management to produce them is a better measure of the program than any attestation.

For sponsors and acquirers

Vulnerability debt is now a diligence line. A target's open-findings count, the age distribution of those findings, its patch cadence against the current volume from its major vendors, and whether it has run an AI-assisted code review are all knowable before signing and are rarely asked for. A software target that sells into banks should be asked what its largest customers have found in its code this year and what it has committed to fix. A services target should be asked how it receives and applies vendor fixes, and how long that takes in practice rather than in policy.

The remediation surge also belongs in the model. Bank of America's number, whatever share of the couple hundred million it turns out to be, is a public data point that AI-scale findings carry real operating cost for at least a quarter or two. A target that has not yet absorbed that cost has not avoided it. It has deferred it, and the buyer will pay it.

Finally, contractor access. If a target or a portfolio company has scaled up outside help for remediation in 2026, the identity verification and access model for those contractors is a post-close integration risk of the first order. The October laptop-farm disclosures listed financial services among the sectors involved. Ask before close, not after.

What we are not saying

We are not saying Bank of America violated any disclosure rule. SEC rules require annual cybersecurity governance disclosure in the 10-K and current reports for material incidents; nothing in the public record suggests a reportable incident, and a voluntary remediation program at this scale relative to the bank's size may well be immaterial. We are not saying any Bank of America contractor is connected to North Korean IT worker schemes or to any other threat; we found no evidence of that and are not suggesting it. We are not saying the Mythos findings at Bank of America were exploited, or that any customer data was exposed. We are not saying the Mythos program and the anti-money-laundering penalties are the same failure; they are different domains with a common structure, and we have said so. We are not saying how the couple hundred million divides, because the bank has not said and we do not know.

We are saying that the second-largest bank in the country disclosed, on a call, that it needed a thousand outside people to keep up with what Mythos and similar AI models found; that it had told shareholders in February that AI would do exactly this; that its filings for the quarter in which it happened say nothing about it, while two Canadian banks named Mythos in their quarterly reports; that its regulators paused some cyber exams of the largest banks to give them time while Bank of America was failing to satisfy a separate order; and that the institutions downstream of it, with none of its resources, are depending on vendors and on a voluntary federal clearinghouse that has not yet published how they can use it. If Bank of America would like to correct any of that, this page will be updated the same day.

Our record of these incidents

• BAE Was Hacked and Every Bankruptcy Court Went Silent, a contractor incident whose consequences land on the banks
• Three Breaches, Three Years: EY, on the gap between what is attested and what is reachable
• Elsevier Was the Name on the Page. RELX Was the Pattern., on a federal warning and a fix that had been public for years
• Instructure's Own Documents Don't Add Up, the document reconciliation KrebsOnSecurity cited
• Shadow AI Agents Are Your New Attack Surface, on AI access nobody inventoried

Sources

• Bank of America Corp., Third-quarter 2026 earnings release, Exhibit 99.1 to Form 8-K, October 14, 2026
• Benzinga, Bank of America Q3 2026 Earnings Call Transcript, October 14, 2026
• Bank of America Corp., Annual Report on Form 10-K for 2025, filed February 25, 2026
• Bank of America Corp., Quarterly Report on Form 10-Q for the period ended June 30, 2026, filed July 31, 2026
• Bank of America Corp., Quarterly Report on Form 10-Q for the period ended March 31, 2026, filed May 1, 2026
• The Bank of Nova Scotia, Second-quarter 2026 report to shareholders, Exhibit 99.1 to Form 6-K, May 27, 2026, and third-quarter report, August 25, 2026
• Canadian Imperial Bank of Commerce, Second-quarter 2026 report, Exhibit 99.1 to Form 6-K, May 28, 2026, and third-quarter report, August 27, 2026
• Connor Jones, The Register, Bank of America had to bring in 1,000 contractors to patch all the bugs Mythos identified, October 15, 2026
• CNBC, Bank of America hired 1,000 contractors to patch software flaws exposed by Anthropic's Mythos, October 14, 2026
• PYMNTS, Bank of America Deploys 4,000 Workers to Patch Vulnerabilities Found by Frontier AI, October 15, 2026
• The Next Web, Bank of America CEO says 1,000 contractors are fixing AI-exposed flaws, October 15, 2026
• Office of the Comptroller of the Currency, OCC Imposes $750 Million Assessment Against Bank of America for Deficient BSA/AML Compliance, NR 2026-89, October 13, 2026
• Axios, Bank of America hit with $1.2 billion penalty over money laundering allegations, October 13, 2026
• Todd Gillespie, Katanga Johnson, Hannah Levitt and Sridhar Natarajan, Bloomberg News via Bloomberg Law, Bessent Urgently Summons Bank CEOs to Discuss Anthropic's New AI, April 10, 2026
• Laura Noonan, Katanga Johnson, Todd Gillespie and Hannah Levitt, Bloomberg News via Bloomberg Law, Wall Street Watchdogs Pause Some Cyber Exams After Mythos Shock, May 19, 2026
• PYMNTS, Banks Slash Patch Times as Anthropic's Mythos Exposes Security Gaps, May 12, 2026 (citing Reuters)
• Constellation Research, Anthropic launches Project Glasswing, Claude Mythos Preview in AI cybersecurity push, April 7, 2026
• Independent Community Bankers of America, ICBA ensures community bank perspective is represented in AI security discussions, April 15, 2026
• Cloud Security Alliance, Executive Order 14409: AI Cybersecurity Deadlines Take Effect, July 6, 2026
• PYMNTS, JPMorgan CEO Warns AI Risks Jumped Tenfold After Mythos, October 6, 2026
• Nupur Anand, Phoebe Seers and Tommy Reggiori Wilkes, Reuters via BNN Bloomberg, Banking watchdogs ramp up AI scrutiny as stability fears mount, sources say, October 13, 2026
• Nextgov/FCW, FBI disrupts North Korean 'laptop farm' used to infiltrate US companies, October 13, 2026
• BleepingComputer, FBI raids DPRK IT worker laptop farms in Texas, US nationals arrested, October 14, 2026
• House Committee on Oversight and Government Reform, Majority Staff, The Equifax Data Breach, December 2018
• The White House, White House Launches Gold Eagle Initiative for Unprecedented Cybersecurity Vulnerability Coordination, July 14, 2026
• Sanjeev Bhasker, Guillermo S. Christensen, Corey Bieber, Tison A. Campbell and Tara P. Ganapathy, K&L Gates, GOLD EAGLE Takes Flight: White House Launches AI-Enabled Cybersecurity Clearinghouse, July 29, 2026
• Microsoft Security Response Center, Security Update Guide CVRF API, monthly documents January 2025 to October 2026, retrieved October 17, 2026
• Cybersecurity and Infrastructure Security Agency, Known Exploited Vulnerabilities Catalog, version 2026.10.16, retrieved October 17, 2026
• U.S. Securities and Exchange Commission, EDGAR full-text search, "Mythos", all form types, April 1 to October 17, 2026, retrieved October 17, 2026