
{{first_name | Reader}},
In partnership with:
Smallstep — SCEP is a password. Passwords get stolen. Real Zero Trust starts with the device — begin with Wi-Fi, extend across apps and infrastructure.
Cite the record - The record behind this brief is public, inspectable, and citable.
The weekly brief is where things get worked out. The daily CISO briefing on Spotify is the fast version: two minutes each weekday on what actually moved. Follow it here.
Thirty Days and Never
In one month, AI was invoked to shorten one disclosure clock and to argue for removing another. The banks have yet to identify a filing that enabled a second attack.
CybersecurityHQ deep dive, 22 November 2026. The SEC filing counts are CHQ's own, taken from the EDGAR full-text index on 22 November and covering filings dated through 20 November. The fix-rate figures are computed from Anthropic's published data as of its 2 November update.
TL;DR
One premise, opposite moves. Citing AI, Project Zero and Anthropic are cutting vulnerability disclosure windows to 30 days, while financial trade groups want the SEC's four-day breach rule rescinded and a White House task force wants enforcement paused where AI overwhelmed defenses.
The banks' security case lacks an example. Their petition names one extortion threat and treats follow-on attacks as a risk, but identifies no SEC filing that enabled a second attack.
The 30-day clock has a steep baseline to beat. Under Anthropic's old 90-day policy, only 33 percent of critical and high findings at least three months old have a recorded patch.
CHQ's view. Keep the breach clock and make it visible by requiring discovery and materiality dates; tie any regulatory relief to how fast a company reports, not to whether the attacker used AI.
On 13 November Google's Project Zero dropped the 30 days of grace it used to give vendors after a fix shipped. It now publishes the details of a vulnerability 90 days after reporting it, whether or not a fix exists. That figure falls to 60 days in February and to 30 in January 2028. Three days later Anthropic set 30 days for the serious flaws its models find, starting in January.
Both gave the same reason. AI finds bugs quickly enough that someone else will find the same one, and it turns a published patch into a working exploit quickly enough that waiting protects nobody.
The day before Google's post, a White House task force applied that fact to a different question. Companies breached by attacks that AI made overwhelming, it said, should not face regulatory enforcement, because the earlier standards "may be outdated".
And since May 2025 five financial trade groups have been asking the SEC to rescind the rule that gives a public company four business days to disclose a breach once it has judged the breach material. In June one of them restated the case for the moment under the title When Disclosure Becomes a Zero Day.
These are different clocks and different risks. A flaw and a breach are not the same kind of secret. But the same premise is doing the work in each argument, and it deserves the same standard of evidence each time.
Whose clock | Before | Adopted or sought | Stated reason |
|---|---|---|---|
A vendor fixing a flaw Project Zero reported | 90 days, plus 30 after a fix | 90 days fixed or not, now; 60 from 2 February 2027; 30 from 3 January 2028 | AI |
A maintainer fixing a flaw Anthropic reported | 90 days | 30 days for high and critical, 45 for the rest, from 1 January 2027, with extensions | AI |
A public company disclosing a material breach | Four business days from the materiality decision | Rescission of the item; disclosure under general principles | AI, among others |
A regulator acting on a breach | Existing standards | No enforcement where AI overwhelmed the defenses; at least 12 months of limited regulatory liability for voluntary reports | AI |
The clock that got shorter
Project Zero's post, by Natalie Silvanovich, makes three claims. Language models let attackers find and exploit bugs faster, so a flaw reported today carries more risk than the same flaw a year ago. More than one party is now likely to find the same bug, so an embargo keeps less of a secret than it did.
And models have cut the time needed to work backwards from a patch to an exploit to nearly nothing, which removes the point of holding details back while a fix rolls out. The team also says it now finds several vulnerabilities a week where it once found a few a year.
Google phased the change over 14 months. Its post says an immediate cut could hurt users by exposing unfixed bugs while vendors were still rebuilding their processes. A Google spokesperson told Dark Reading that every 2026 report the team made against a volunteer-maintained project was fixed within 30 days, with one exception that took 31. Flaws already exploited in the wild keep their seven-day window.
Anthropic's policy starts sooner and has more moving parts. From 1 January, high and critical findings get 30 days and everything else gets 45, down from 90. A maintainer who is engaged and making progress can ask for 14 more days. The company says it will not send a project large volumes of findings without first agreeing a pace the maintainer can sustain.
By default, the company's ledger reveals a finding's identifier, project and class of bug when its window closes, sometimes with Anthropic's analysis, and full technical detail generally follows 30 days after a patch is available. Those are stated defaults, with exceptions in both directions. The policy does not say when full detail follows for a flaw still unpatched at the deadline. Reports are meant to reflect findings a human researcher has reviewed.
The objection from people who run disclosure programs is about capacity. Katie Moussouris told Dark Reading that a deadline does not produce a fix and that AI companies are sending maintainers reports without helping to fix them. Marcus Hutchins said maintainers are already patching as fast as they can. Dustin Childs of the Zero Day Initiative said he already sees incomplete patches at 120 days and fears more of them at 30.
Anthropic publishes enough data to set a baseline. Its dashboard shows 5,087 findings reported to 680 open source projects, of which 629 have a patch that landed after the report. That 12 percent mixes ages and severities, so we cut it by both.
Findings reported, by the severity Claude assigned | Reported | Patched after the report | Share |
|---|---|---|---|
Critical and high, all ages | 1,813 | 500 | 28% |
Critical and high, discovered at least 90 days before 2 November | 1,457 | 479 | 33% |
Medium and low, all ages | 3,274 | 129 | 4% |
Among serious findings at least three months old, one in three has a recorded fix. Three limits apply. The published date is the date of discovery, not of the report, so the true age since a maintainer was told is shorter. A mitigation that is not an upstream patch does not count.
And 3,225 of the 5,087 were sent to maintainers directly, without the external-firm review the other 1,862 received; the dashboard warns that group may contain false positives. The published data does not give a fix rate for the reviewed findings alone. Of the 237 findings Anthropic has publicly revealed, 217 are marked fixed.
All of this describes reports made under the old 90-day policy. The new one applies to findings reported from 1 January and has not started. The table is a baseline for judging it.
The clock that may be removed
The SEC rule has applied since December 2023. A public company that determines a cybersecurity incident is material must file a Form 8-K under Item 1.05 within four business days of that determination.
The rule's first instruction says the determination must be made without unreasonable delay after discovery. Its fourth says the company need not disclose technical detail about its systems, its vulnerabilities or its response. The Attorney General can delay a filing that would endanger national security or public safety.
In May 2024 the SEC's corporation finance director added staff guidance. Item 1.05 is for incidents a company has determined to be material. An incident it has not yet assessed, or has judged immaterial, belongs under the general-purpose Item 8.01. Filing under Item 8.01 does not remove the duty to decide materiality without unreasonable delay, and a company that later finds the incident material must then file under Item 1.05.
On 22 May 2025 SIFMA, the American Bankers Association, the Bank Policy Institute, the Independent Community Bankers of America and the Institute of International Bankers petitioned the SEC to rescind Item 1.05.
In January 2026 Chairman Paul Atkins opened Regulation S-K to comment, and the same five groups wrote in April asking for Item 1.05 and the annual cyber governance disclosure to go. The wireless industry's trade group filed a similar letter on 2 October.
American Banker reported in June that Atkins was reviewing the rule and that a commissioner who dissented from it in 2023 now sits in the majority. We found no proposed rule as of 22 November.
Several of the petition's arguments have nothing to do with attackers: that the rule conflicts with confidential reporting to other agencies, that the Attorney General delay is too narrow to use, that the filings tell investors little, and that compliance is a burden.
On security, its evidence is of three kinds. It names one case: in November 2023 the ransomware group AlphV reported its own victim, MeridianLink, to the SEC for not disclosing, a month before the rule applied. The petition presents that as the rule being turned into extortion leverage.
It says it knows of other threats that cited the rule, without naming them. And it argues that a company disclosing an ongoing incident may draw new attacks on a weakness it has not yet closed. It presents that last point as a risk. It cites no filing after which it happened.
The Bank Policy Institute's June post added the AI argument: frontier models could combine filings, press coverage and technical indicators to refine an attack in near real time. Its examples are from Log4j, a vulnerability disclosure.
What the filings show, and what they do not
CHQ searched the EDGAR full-text index on 22 November for every 8-K whose declared item list includes 1.05.
Period | Original 8-Ks under Item 1.05 | Amendments |
|---|---|---|
18 to 31 December 2023 | 2 | 1 |
2024 | 24 | 14 |
2025 | 15 | 4 |
2026, to 20 November | 18, from 17 companies | 7 |
Our method reproduces the petitioners' own figures. Their petition counts 17 Item 1.05 filings before the staff guidance of May 2024 (the two from December 2023 and 15 from 2024) and nine in the rest of that year. Our index search returns 17 and nine for the same periods. That checks the record to the end of 2024. Nobody outside CHQ has yet reproduced 2025 or 2026.
The 2026 count is 18 original filings in nearly eleven months, and none since 13 October.
We then looked for incident disclosures made elsewhere. In 2026, 26 8-Ks from 21 companies were filed under Item 8.01 alone with the phrase "cybersecurity incident" in the main document. We have not read them, so this is a count of filings that use the phrase, and some will concern a supplier's incident or a past one. A filing that describes ransomware or data theft in other words is not among the 26; the appendix gives a wider count.
Two of the 21 companies went on to Item 1.05, which is the path the staff guidance describes: Boston Scientific 13 days after its Item 8.01 filing, and Stryker 29 days after, by amendment. The other 19 have no Item 1.05 filing in the index this year.
One of the 19 shows what an outside reader cannot see. Hasbro filed under Item 8.01 on 1 April. Cybersecurity Dive later reported that the company expected $20 million in remediation costs and $40 million to $60 million in delayed revenue.
No Item 1.05 filing from Hasbro appears in the index through 20 November. That may be a sound determination that the incident was not material to a company of its size. The filings do not say whether a determination was made, or when.
What the counts establish. Our search identified 59 original Item 1.05 filings in just under three years. That is the population of original filings against which the petitioners' claim can be tested. Amendments, and anything companies said through other channels, sit outside it. And the petitioners' own description of the shift to Item 8.01 is accurate.
What they do not establish. A low count does not show under-reporting. The right denominator is incidents that met the materiality standard, and that number is not public.
Nor does our reading of filings settle what an attacker could learn. The three incident filings CHQ has recorded in full this year, from Boston Scientific, Astrana Health and SBC Medical Group Holdings, give dates and impact and no method of entry. Three filings cannot prove that no filing's timing or confirmation of compromise ever helped an adversary.
The grace year
The task force met for the second time on Thursday 12 November. By Cybersecurity Dive's account of its statement, it made two recommendations.
The first is that regulators should not take enforcement action against companies hit by cyberattacks that result from AI overwhelming their defenses. The second is that federal agencies should limit regulatory liability for good-faith voluntary reporting of vulnerabilities and incidents, for at least 12 months, where that is lawful, practical and in the public interest.
Both concern regulatory enforcement. The report says nothing about investor disclosure or private lawsuits, and it does not say which regulators are covered, how good faith would be judged, or who decides that an attack was AI-driven. We have not read the statement itself.
That last gap matters because the label is already applied on inference. In a case CHQ recorded this week, Rapid7 published a threat report on an intrusion through Apache Axis and noted that the attacker's commands wrapped their output in labelled delimiters with sequential task numbers. It read that structure as a sign of AI-assisted tooling.
That is a reasonable inference in ordinary threat reporting. Nobody in that case has asked a regulator for relief, and a regulator would presumably ask for more. It shows what the phrase rests on today.
CHQ's view
The premise is right. Since August CHQ has counted seven cases, at seven vendors, in which exploitation began before the vendor's first advisory. The gap ran from one day to 61. In those cases silence protected no defender, because an attacker already had the flaw.
A breach is a different secret, and the banks' strongest point should be taken seriously. The intruder knows about its own intrusion. Other attackers may not, and public confirmation of a live compromise can tell them where to look and tell the intruder it has been seen. That is a real security concern, and the rule's instruction on technical detail and its Attorney General delay exist because of it.
But it is an empirical claim, and the petitioners have not yet put a case behind it. Fifty-nine original filings are on the public record. The petitioners name one incident, and it concerns a criminal using the rule's existence as leverage. They cite no filing that was followed by a second attack.
If AI makes information more dangerous in an attacker's hands, the question for anyone who wants less disclosure is what evidence justifies keeping a material incident from investors, customers and other defenders.
We have not shown that no such case exists. We have shown that the people best placed to find one did not cite it. Until they do, the security argument for rescission rests on a risk the rule already has two tools to manage. The petition's other arguments, on duplication, burden and usefulness to investors, stand or fall separately and are not answered here.
Make the trigger visible before changing it. The rule already requires a materiality decision without unreasonable delay. Public filings do not consistently provide the dates needed to assess that. A filing does not have to state when the incident was discovered or when the determination was made, and an Item 8.01 filing does not have to say whether one has been made at all.
Requiring those two dates in every Item 1.05 filing, an explanation of any long gap between them, and a statement of status in every Item 8.01 incident filing, would let anyone measure the interval. Two dates do not prove a determination was timely. They show how long it took, and the explanation gives regulators and investors something to weigh, with no technical detail required.
A discovery date is not free of risk: during a live intrusion it tells the intruder when it was seen. The existing delay for national security and public safety could be widened to cover that case. Our judgment is that the gain in accountability outweighs the added risk.
"AI overwhelmed us" cannot be a standard for relief. The finding would be written by a response firm retained by the company seeking the relief. Every incident report would have a reason to contain the phrase. The task force's second recommendation is the sound one: limit regulatory liability for companies that report quickly and in good faith. That should depend on how fast and how fully a company reported. The attacker's tooling should have nothing to do with it.
The fix data is a baseline the short clock has to beat. We do not object to the direction. Google's phase-in is defensible on the record it describes, and Anthropic's policy has extensions and a pacing commitment.
But under the old 90-day policy, on Anthropic's own published figures, a third of serious findings at least three months old have a recorded fix. The new policy has not started, so that is a starting point and no verdict.
From January the test is what share of confirmed high and critical findings are fixed or effectively mitigated within 30 days of the maintainer receiving the report. That is a different measure from ours, which counts a fix that arrived on day 70. The first cohorts mature in February, and it will take several months of them to say anything reliable.
The published data carries no report date today, so the test needs Anthropic to add one. If the share turns out low, the clock will be moving risk from the finder's ledger to the operator's network.
Operators are missing from the public record of all three processes. Google says it consulted vendors. The petitions are from banks and carriers. The task force is working with AI labs. We found no operator group in the sources we read. That may be a gap in the record, and it is worth asking about.
What cuts against it
Nothing has been decided. The SEC has proposed no change, and the task force issued recommendations, which we read only through one outlet's report.
Liability limits for voluntary reporting have precedent. The 2015 information-sharing law works on that same bargain, and the task force supports renewing it.
We did not independently search for a harmful filing. Our claim is only that the petitioners cite none; a case may exist, and their petition raises non-security objections we do not address.
The fix-rate baseline is imperfect. It uses discovery dates, Claude's severity ratings, and data from the old 90-day policy; it cannot show how maintainers will respond to a 30-day deadline.
One Google data point is unverified. The 30-day fix record for volunteer projects was given to a reporter and is not in Project Zero's published post.
What to do with this
Plan for public notice of unpatched flaws. By January 2028 Project Zero and Anthropic will both be on 30 days, and others are expected to follow. Know which internet-facing software you run that is maintained by volunteers, and decide now what you switch off or wall off when a flaw in it is announced without a fix.
Ask vendors for their fix time, in days. The median time from an external report to a shipped fix now decides whether you meet one of these deadlines unpatched. A vendor that cannot state it does not measure it.
Do not rely on SEC filings to learn of a supplier's breach. Item 1.05 covers incidents a company has judged material, and it produced 18 original filings this year. It was never meant to report every supplier compromise. A contractual notification duty, with a deadline measured from discovery, reaches further and sooner.
If you file, consider stating the two dates. The date of discovery and the date of the materiality determination let a reader see how the rule was followed. Weigh that against what a discovery date tells an intruder who may still be inside, and decide the template before the incident.
Treat "AI-driven" as a claim. When a report says an attack was AI-assisted, ask what the evidence is. If relief from enforcement ever depends on that phrase, the standard of proof will matter a great deal.
