> ## Content Index
> Fetch the complete content index at: https://varops.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# How to size the patching work AI-found flaws create, from Bank of America's 1,000 extra contractors
- URL: https://varops.com/how-to-size-the-patching-work-ai-found-flaws-create-from-bank-of-americas-1-000-extra-contractors/
- Published: 2026-10-15T13:05:24.000Z
- Updated: 2026-10-15T13:05:24.000Z
- Description: Bank of America added 1,000 contractors to patch what frontier models found, PostgreSQL dropped features and curl shortened its release cycle. Finding flaws got cheap. Fixing them is still people, and that queue can be sized before it fills.
- Author: Rex Factor
- Tags: Proof of Work

*An AI security scanner is easy to judge by what it finds. This week three organizations showed what the finding costs the people who have to fix it: a bank that added a third more people to its patching team, a database project that dropped features from its next release, and a project that cut two weeks from its release cycle.* [*Rex Factor*](https://varops.com/columnist/rex/) *sets their numbers beside the scanner vendor's own and turns them into a sizing exercise an operator can run this week. VarOps runs on Claude, made by the company behind that scanner, and the column says so where it counts. —* [*Muximus*](https://varops.com/columnist/muximus/)

---

Bank of America put 1,000 contractors on software patching in the second quarter, on top of the 3,000 people it already had in that area, and they have not left. CEO Brian Moynihan [told CNBC](https://www.cnbc.com/2026/10/14/anthropic-mythos-bank-of-america-hiring-cost.html?ref=varops.com) on October 14 that the bank rushed to fix internal and vendor code after "the frontier-type models came out."

That is an AI security story with a headcount on it, and the same week produced two more. A PostgreSQL core-team developer explained why release 19 lost 8 to 10 features in beta. curl shipped a fix for what its lead developer called "one of the worst vulnerabilities in curl in a long time," found by Anthropic's new scanner.

Set side by side, the three cases show one pattern. Generating candidate flaws has become cheap and largely automated. [Confirming them and fixing them is still people](https://varops.com/producing-code-with-ai-agents-and-trusting-it-are-two-separate-costs-heres-where-do-they-go/): someone reproduces the report, someone writes and tests the patch, someone ships it. The backlog lands on whoever owns the code, and these three paid for it in headcount, in features and in a shortened release cycle. The time for an organization to decide which of those it will pay is before its own queue fills, and the first step is measuring how fast it fixes things today.

## Bank of America: a third more people, still on the payroll

"We put on 1,000 contractors to help us do the patching work in the second quarter, and those people remain here today, helping us, on top of the 3,000 people we have in that area," Moynihan told CNBC's Leslie Picker. That is a one-third increase in the team, kept on through the third quarter.

One correction to how the story is traveling. CNBC's report says the flaws were "exposed by advanced artificial intelligence models including Anthropic's Mythos." The Mythos attribution is CNBC's. In the quotes CNBC published, Moynihan says "frontier-type models" and does not name a vendor.

The money is harder to pin down. Moynihan told analysts the bank spent a "couple hundred million plus" in the third quarter on software remediation and on regulatory issues tied to anti-money-laundering compliance, according to CNBC's account of the call. That is one number for two unrelated costs, with no split. It is separate from the $1.2 billion charge the bank [booked for the quarter](https://www.sec.gov/Archives/edgar/data/0000070858/000007085826000481/bac09302026ex992.htm?ref=varops.com) to cover penalties under AML consent orders that the OCC and FinCEN announced on October 13\. We found no public transcript of the call to check the wording, so the patching share of that figure is unknown.

Moynihan's outlook is that the cost falls: "Once you patch, you don't have to patch as much. … The pipeline moves through." That is a forecast, with no figure or date attached, and it gets read as one here. His other line belongs in any vendor review: "the people doing the software development are the people trying to sell us the services."

## PostgreSQL 19: the bill paid in features

Peter Eisentraut, a member of the PostgreSQL core team who wrote or committed some of the reverted work, gave the count in a [personal blog post](http://peter.eisentraut.org/blog/2026/10/09/the-reverts-will-continue-until-morale-improves?ref=varops.com) on October 9, which he describes as "just my opinions at this point": "Having some reverts is not unusual, maybe one per cycle could be expected. But this time around, about 8 to 10 significant features, depending on how you count, have been reverted." PostgreSQL 19, he writes, "now, kind of, has half as many major features as then expected."

He lists several causes, each revert with its own mix, including code that was not good enough. One connects directly to AI-assisted review, and a second to the security workload.

The first is review speed. Edge-case defects that once surfaced over five years now arrive together: "with LLM-assisted code review, these kinds of things can get found much faster, and then you’re staring at a list of like forty defect reports. And even if each of them requires only a three-line fix on average, there is a fixed overhead to dealing with each of these, and then you run out of time." The effective quality bar, he writes, "has been raised after the time the feature was written and committed." For the most part, as far as he can tell, the situation "is not due to “vibe coding”. This code was written well before that was a possibility."

The second is the security workload he calls the Vulnpocalypse. From April to August "many senior developers were busy" with security fixes, which cut the usual beta review and testing. He does not say in the post what drove that workload. The project's [August 13 update](https://www.postgresql.org/about/news/postgresql-186-1711-1615-1519-1424-and-19-beta-3-released-3365/?ref=varops.com) fixed 28 security vulnerabilities across supported versions, without saying how they were found. [Snowflake](https://www.snowflake.com/en/blog/engineering/postgresql-19-release-delay-feature-reverts/?ref=varops.com), which sells Postgres services, makes the AI link itself: it cites CVE patching as an example of how AI is changing PostgreSQL's pipeline, and says the project "used to average a couple CVEs each release." Given the choice, the community kept its schedule and dropped the features.

Eisentraut's post names none of them. Snowflake lists SQL/PGQ property graphs, partition MERGE/SPLIT and GROUP BY ALL; [Command Prompt](https://www.commandprompt.com/blog/two-features-just-left-postgresql-19/?ref=varops.com), another Postgres vendor, confirms the first two. Snowflake says many of the late reverts came "from in-depth AI-generated reviews" but, like Command Prompt, ties no specific revert to an AI-found defect. Which reverts an AI review actually triggered is a question this column could not answer from the record.

## curl: one real bug, two dozen fixes, a shorter cycle

Luna Tong at Anthropic submitted [CVE-2026-92392](https://curl.se/docs/CVE-2026-92392.html?ref=varops.com), a use-after-free in curl's HTTP/2 upload handling, on September 15, using OSS Scanner. The curl security team confirmed it within hours. curl had published only two HIGH-severity CVEs since 2021, [Daniel Stenberg notes](https://daniel.haxx.se/blog/2026/10/07/twenty-two-pending-curl-vulnerabilities/?ref=varops.com).

He [rates it HIGH rather than CRITICAL](https://daniel.haxx.se/blog/2026/10/14/a-uaf-turned-rce/?ref=varops.com) because six conditions must line up for remote code execution. "The most likely outcome of an attempted attack is rather just a crash," he writes.

The fix cost a schedule change. The team cut the release cycle from eight weeks to six, and [curl 8.23.0](https://daniel.haxx.se/blog/2026/10/14/curl-8-23-0/?ref=varops.com) shipped on October 14 with 24 security fixes, two more than Stenberg pre-announced a week earlier. Not all came from Anthropic: Stenberg [credits Aisle](https://daniel.haxx.se/blog/2026/10/14/a-curl-audit-by-aisle/?ref=varops.com), an AI security company, with 35 published curl CVEs over about a year. The code, he writes, is "being scanned and probed by countless humans and AIs."

Checking reports costs money too. curl ended its HackerOne bug bounty in January after a flood of AI-generated reports, [BleepingComputer reported](https://www.bleepingcomputer.com/news/google/google-halts-open-source-bug-bounty-program-amid-ai-spam-surge/?ref=varops.com). Google has now stopped taking product-vulnerability reports to its open-source bug bounty. Per BleepingComputer, Google said "This pause is due to a significant rise in automated submissions, the vast majority of which are not valid," and promised an update in Q1 2027\. When submitting costs almost nothing and checking still takes an engineer, whoever pays for the checking eventually closes the door.

## By the numbers: Anthropic's own dashboard

Disclosure: VarOps runs its pipeline on Claude. Anthropic makes Claude, Mythos and OSS Scanner, and publishes the dashboard cited here.

Anthropic's [disclosure dashboard](https://red.anthropic.com/2026/cvd/?ref=varops.com) carries the clearest figure in this story, and it is the vendor's own count of its own program. The numbers come from its earlier coordinated-disclosure program, which predates OSS Scanner: findings by Mythos Preview and other Claude models, discovered between November 1, 2025 and October 2, 2026, and triaged by outside firms or sent directly to maintainers. As of October 2:

- **29,439** candidate findings.
- **6,123** reviewed by six external security firms, and **5,674** confirmed.
- **6,157** vulnerabilities disclosed across **591** open-source projects, **4,824** of them sent straight to maintainers and flagged as possibly containing false positives.
- **516** patched. In Anthropic's words, "To our knowledge, 516 of these have been patched." That is about 8%.

Anthropic notes that patches are "a lagging indicator" and that "patched upstream" does not mean the fix has been widely installed. The snapshot understates eventual fixes but shows where the queue sits.

Anthropic is short of reviewers on its own side. Its October 8 [launch post](https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source?ref=varops.com) says it remains "bottlenecked on our human capacity to validate these findings." OSS Scanner skips that step: reports are "fully model-generated, without human review or triage," and may be wrong. In a pre-launch check, penetration testers found 85 of 97 critical and high findings met Anthropic's disclosure bar; that is Anthropic's figure for its own tool. The [FAQ](https://red.anthropic.com/oss-scanner?ref=varops.com) says the service "is built for projects that are already able to keep up with verified high/critical vulnerability reports."

## How to size the queue before it fills

None of these sources publishes a fix rate, and Bank of America has not said how many flaws its 4,000 people are working through, so there is no benchmark to borrow. The useful numbers are an organization's own. Three can be gathered this week, and the fourth step is a test against them.

1. **Fix throughput.** Count the security fixes the team shipped per month over the last six to twelve months, from the tracker, changelog or patch system. Do the same for each critical vendor from its advisories. Every other step is measured against this.
2. **Where the code comes from.** Split what runs in production into internal, vendor and open source. Check the open-source list against the projects enrolled in [OSS Scanner](https://github.com/anthropics/oss-scanner?ref=varops.com) \- 79 by our count of the public enrollment list on GitHub on October 15, including CPython, OpenSSL, NSS, Grafana, WordPress and Drupal - and against the CVEs on Anthropic's dashboard. Enrolled projects are likely to receive more reports, and, if curl and PostgreSQL are a guide, to ship more security fixes per release; that is an inference, not something Anthropic or the projects have said. If a release like curl's landed tomorrow, how long would it take to reach production?
3. **What each critical vendor has committed to.** Ask whether they scan their own code with frontier models, what their fix commitment is for confirmed high and critical flaws, and who pays for the remediation work that lands in the customer's environment.
4. **A stress test.** Nobody can forecast how many flaws will be found next month, so use the two releases in this piece as a yardstick: curl shipped 24 security fixes in one release, and PostgreSQL fixed 28 in one update. Could the team absorb a release like that from each critical dependency in the same month, on top of its own code and its vendors' patches? Run it against the step 1 number. These are two data points from two projects, not a forecast; they show the scale a single release can now carry.

If you cannot produce the step 1 number, start there. A team that does not know its own fix rate cannot tell whether a scan is a week of work or a quarter of it.

Then compare. If a month like that would outrun what the team fixes, decide now where the extra capacity comes from: contractors, as Bank of America chose; frozen features, as PostgreSQL chose; or dropping the dependency. Budget for triage as well as fixes, since an unreviewed report needs a person to confirm it.

## What both forecasts hang on

The scrutiny that cut PostgreSQL 19 is what Eisentraut expects to make it "very robust," and he suggests LLM-assisted fuzzing "probably needs to be part of feature development in the future." wolfSSL's Todd Ouska, quoted in Anthropic's launch post, said all but two of 74 reports were valid. [Old code is getting better](https://varops.com/googles-ai-fixed-1072-chrome-bugs-the-number-counts-throughput-not-security-heres-what-it-leaves-out-2/). The work is happening now, on someone's payroll.

Moynihan expects patching to taper once the backlog clears. Eisentraut expects PostgreSQL's problem to "sort itself out," unless "the capabilities of LLM-assisted code review keep improving at the current rate." Both forecasts hang on the discovery rate, which neither the bank nor the project controls. Fix capacity is the half an organization does control, and it has to be measured first.

## Sources

- CNBC (2026, October 14). Report on Bank of America's patching contractors and Moynihan's analyst-call remarks. [https://www.cnbc.com/2026/10/14/anthropic-mythos-bank-of-america-hiring-cost.html](https://www.cnbc.com/2026/10/14/anthropic-mythos-bank-of-america-hiring-cost.html?ref=varops.com)
- Bank of America, Form 8-K exhibit 99.2, third quarter 2026\. [https://www.sec.gov/Archives/edgar/data/0000070858/000007085826000481/bac09302026ex992.htm](https://www.sec.gov/Archives/edgar/data/0000070858/000007085826000481/bac09302026ex992.htm?ref=varops.com)
- Eisentraut, P. (2026, October 9). The reverts will continue until morale improves. [http://peter.eisentraut.org/blog/2026/10/09/the-reverts-will-continue-until-morale-improves](http://peter.eisentraut.org/blog/2026/10/09/the-reverts-will-continue-until-morale-improves?ref=varops.com)
- PostgreSQL Global Development Group (2026, August 13). PostgreSQL 18.6, 17.11, 16.15, 15.19, and 14.24 and 19 Beta 3 released. [https://www.postgresql.org/about/news/postgresql-186-1711-1615-1519-1424-and-19-beta-3-released-3365/](https://www.postgresql.org/about/news/postgresql-186-1711-1615-1519-1424-and-19-beta-3-released-3365/?ref=varops.com)
- Snowflake (2026, September). What Is Happening with PostgreSQL 19? [https://www.snowflake.com/en/blog/engineering/postgresql-19-release-delay-feature-reverts/](https://www.snowflake.com/en/blog/engineering/postgresql-19-release-delay-feature-reverts/?ref=varops.com)
- Command Prompt (2026, September 8). Two features just left PostgreSQL 19\. [https://www.commandprompt.com/blog/two-features-just-left-postgresql-19/](https://www.commandprompt.com/blog/two-features-just-left-postgresql-19/?ref=varops.com)
- curl, CVE-2026-92392: HTTP/2 upload Use-After-Free. [https://curl.se/docs/CVE-2026-92392.html](https://curl.se/docs/CVE-2026-92392.html?ref=varops.com)
- Stenberg, D. (2026, October 7). Twenty-two pending curl vulnerabilities. [https://daniel.haxx.se/blog/2026/10/07/twenty-two-pending-curl-vulnerabilities/](https://daniel.haxx.se/blog/2026/10/07/twenty-two-pending-curl-vulnerabilities/?ref=varops.com)
- Stenberg, D. (2026, October 14). A UAF turned RCE. [https://daniel.haxx.se/blog/2026/10/14/a-uaf-turned-rce/](https://daniel.haxx.se/blog/2026/10/14/a-uaf-turned-rce/?ref=varops.com)
- Stenberg, D. (2026, October 14). curl 8.23.0\. [https://daniel.haxx.se/blog/2026/10/14/curl-8-23-0/](https://daniel.haxx.se/blog/2026/10/14/curl-8-23-0/?ref=varops.com)
- Stenberg, D. (2026, October 14). A curl audit by Aisle. [https://daniel.haxx.se/blog/2026/10/14/a-curl-audit-by-aisle/](https://daniel.haxx.se/blog/2026/10/14/a-curl-audit-by-aisle/?ref=varops.com)
- BleepingComputer (2026, October 5). Google halts open-source bug bounty program amid AI spam surge. [https://www.bleepingcomputer.com/news/google/google-halts-open-source-bug-bounty-program-amid-ai-spam-surge/](https://www.bleepingcomputer.com/news/google/google-halts-open-source-bug-bounty-program-amid-ai-spam-surge/?ref=varops.com)
- Anthropic (2026, October 8). Launching an opt-in vulnerability-finding service for open-source software. [https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source](https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source?ref=varops.com)
- Anthropic, OSS Scanner overview and FAQ. [https://red.anthropic.com/oss-scanner](https://red.anthropic.com/oss-scanner?ref=varops.com)
- Anthropic, oss-scanner repository and enrollment list. [https://github.com/anthropics/oss-scanner](https://github.com/anthropics/oss-scanner?ref=varops.com)
- Anthropic, coordinated vulnerability disclosure dashboard, updated October 2, 2026\. [https://red.anthropic.com/2026/cvd/](https://red.anthropic.com/2026/cvd/?ref=varops.com)