What is patch diffing — why a fix is a map, 63 days to −7
Patch diffing is an analysis technique that compares software before and after a security patch, finds what changed, and works backward to where the fixed vulnerability was. A patch shows what it repairs, so the moment it ships it is also a clue for anyone targeting systems that have not updated yet. Mandiant's average time-to-exploit fell from 63 days in 2018-2019 to 5 days in 2023, and to an estimated minus 7 days in 2025, with attacks that came before the patch pulling the average below zero. On November 16, 2026, Google said it had found exploit_agent on a real attacker server, the first framework it has seen that compares code before and after a patch, writes an exploit and runs it with no human involved. Defenders use the same technique to verify fixes and write detection rules. For individuals the answer is automatic updates. For organizations it is ordering patches by confirmed exploitation (KEV) and exploitation probability (EPSS) and applying them within days.
The three lines
- Meaning — comparing pre-patch and post-patch versions to locate and characterize the vulnerability that was fixed; attackers and defenders use the same tools
- Speed — Mandiant's average time-to-exploit went from 63 days (2018-2019) to 5 days (2023) to an estimated −7 days (2025)
- AI — on November 16, 2026, Google reported exploit_agent, the first framework it has seen turning patches into exploits with no human in the loop
Key questions
- what is patch diffing
- **An analysis technique that compares a program before and after a security patch, isolates the changed code, and infers from that change what the vulnerability was. When done on compiled programs it is also called binary diffing.** | Item | Detail | |---|---| | What is compared | The version just before the patch and the version just after (source code or executables) | | What it reveals | Which function gained or changed which check, meaning the location and type of the flaw | | Main tools | BinDiff (maintained by Google, Apache 2.0), Diaphora (open-source IDA plugin) | | Defensive uses | Verifying a fix is complete, writing detection rules, spotting flaws fixed without an advisory | | Offensive use | Building n-day exploits for systems that have not updated | | How to stop it | The technique cannot be stopped. Applying the patch quickly is the only way out of the target set |
- zero-day vs n-day vulnerability
- **A zero-day is exploited before a patch exists. An n-day (or one-day) is exploited after the patch is out, on systems that have not applied it. Patch diffing is how n-day exploits get made.** | | Zero-day | N-day (one-day) | |---|---|---| | Patch | None | Available | | What the attacker starts from | Has to find the flaw alone | The patch and advisory are clues | | The user's defense | Effectively none | Updating | | Share of exploited flaws in Mandiant's 2023 data | 97 of 138 (70%) | 41 of 138 (30%) | | Who has to act | The vendor | Users and administrators |
- how long after a patch is a vulnerability exploited
- **It depends on who is counting and how, but every series points the same way. Mandiant's average went from 63 days to 5 and then below zero, and among flaws exploited only after a patch, more than half were hit within a month in 2023.** | Source | Measure | Figure | |---|---|---| | Mandiant | Average time-to-exploit | 63 days (2018-2019) → 44 → 32 → 5 days (2023) → estimated −7 days (2025) | | Mandiant | 41 n-days from 2023 | 12% within a day, 29% within a week, 56% within a month | | Rapid7 | Median time from publication to CISA KEV listing, high and critical flaws | 8.5 days → 5.0 days | | VulnCheck | Share with exploitation evidence on or before the CVE publication date | 28.93% in 2025, 23.43% in the first half of 2026 | | VulnCheck | Median time from CVE publication to exploited-list entry | 120 days (2025) → 80 days (first half of 2026) |
Patch diffing is the technique of placing software side by side as it was before and after a security patch, finding what changed, and working backward to where the fixed vulnerability was. A patch is a cure, and it is also a signpost pointing at the spot that was broken. A vendor can write a one-line advisory saying it fixed a remote code execution flaw, and a comparison of the code before and after will still show which function was missing which check. For users who have installed the update, the matter is closed. Every system that has not installed it now carries a hole whose location is known.
Google Threat Intelligence Group (GTIG) said in a quarterly report published on November 16, 2026, that it had for the first time found a framework doing this work with no human involved, on a real attacker's server. The tool, which GTIG calls exploit_agent, was set up to keep comparing the code of popular web application software before and after patches, infer the vulnerability, write an exploit and run it. The details of that discovery are in our separate November 18, 2026 report on Google's exploit_agent finding. This article covers the principle underneath: what patch diffing is, why it cannot be prevented, how far the gap between patch and exploitation has shrunk in published statistics, and what individuals and organizations should do now that the gap is short. It does not describe how to attack any specific vulnerability.
1. What patch diffing is: the repaired spot is the broken spot
Diffing means extracting the differences between two versions. Think of tracked changes in a word processor. Compare the old and new drafts of a contract and you see which clause changed, and from the changed clause you can guess what the other side was worried about. Software patches work the same way.
Security patches are usually small. One line that checks a length, one condition that checks a permission, one step that clears an address after memory is released. So if you compare before and after and find the newly added check, you can reason backward to what was possible without it. A new length check means that data could previously be written past the end. A new permission check means the function could previously be reached without permission. For a use-after-free bug, the kind covered in our November 14, 2026 explainer, the clue is the spot where code was added to clear a pointer to released memory.
The analysis generally runs in the following order. This is a conceptual outline, and attackers and defenders take the first four steps identically.
| Step | What is done | What makes it hard |
|---|---|---|
| 1. Get both versions | Obtain the build just before the patch and the one just after | In open source the fixing commit is public as is. Commercial products ship executables only |
| 2. Extract differences | Pair up functions and keep only those that changed | A different compiler setting can make all the machine code look different, which is why tools are needed |
| 3. Filter noise | Separate security fixes from feature work and refactoring | A monthly cumulative update mixes hundreds of fixes |
| 4. Understand the cause | Work out what the added check prevents and what input reaches that point | The step that takes the most expertise |
| 5. Fork | Defenders go on to detection rules and patch verification; attackers go on to exploit development | The same knowledge is used in opposite directions |
In projects with public source code, step 2 is easy. The repository's change history is the difference. GTIG wrote that attackers also draw on source code from pre-release or preview channels, release notes, security research and proof-of-concept code as comparison material. Commercial software distributed only as executables is analyzed through binary diffing: the executable is disassembled, each function's call relationships and control flow are turned into graphs, and the graphs of the two versions are matched against each other.
Two tools dominate.
| Tool | Maintainer and license | Characteristics | Use stated in its documentation |
|---|---|---|---|
| BinDiff | Google, Apache 2.0 | Works with IDA Pro, Binary Ninja and Ghidra; supports x86, ARM, MIPS, PowerPC and others | Finding identical and similar functions across binaries, and isolating security fixes in vendor patches |
| Diaphora | Joxean Koret, AGPL 3.0 | IDA plugin that compares assembly, control-flow graphs and pseudo-code | Lists detection of potentially fixed vulnerabilities in patch diffing as a feature unique to it |
Both tools are public, and both name patch analysis as a purpose. The BinDiff README lists isolating vulnerability fixes in vendor patches among its uses. The Diaphora README shows screenshots of it identifying the patched functions in past Windows fixes. This is standard practice in the security industry, not a hidden craft. And that is the point to hold on to: patch diffing is not something that can be banned or blocked. To deliver a patch, a vendor has to send the changed code to users, and anyone who receives code can compare it.
2. From patch to exploitation: what the statistics show
A vulnerability that is exploited after its patch is public, on systems that have not applied it, is called an n-day, meaning N days have passed since disclosure. One exploited soon after disclosure is often called a one-day. A zero-day, the subject of our August 10, 2026 explainer, is exploited before a patch exists. Mandiant defines zero-days as vulnerabilities exploited before patches are made available, and n-days as vulnerabilities first exploited after patches are available.
| Zero-day | N-day (one-day) | |
|---|---|---|
| Patch | None | Available |
| Attacker's effort | Has to find the flaw unaided | Reads the patch and advisory to narrow the search |
| Typical technique | Fuzzing, code auditing | Patch diffing |
| Who is exposed | Everyone, including fully updated systems | Systems that have not updated |
| The user's defense | Effectively none | Updating |
| Share of exploited flaws in 2023 (Mandiant) | 97 (70%) | 41 (30%) |
The best-known measure of the gap between patch and exploitation is Mandiant's time-to-exploit (TTE). It sets the day a patch is released as zero and averages the number of days to first exploitation. When exploitation comes before the patch, the value is negative.
| Period covered | Average time-to-exploit | Source |
|---|---|---|
| 2018-2019 | 63 days | Mandiant's 2023 analysis (published October 16, 2024) |
| 2020 to early 2021 | 44 days | Same |
| 2021-2022 | 32 days | Same |
| 2023 | 5 days | Same; 138 vulnerabilities analyzed, 15 outliers excluded |
| 2025 | Estimated −7 days | M-Trends 2026 (published March 24, 2026) |
Two months became five days in five years, and the figure then passed zero. In its summary of M-Trends 2026, Mandiant read minus 7 days as meaning exploitation is routinely occurring before a patch is released.
Two cautions apply to this table. First, the average includes zero-days. A zero-day is negative by definition, so the average falls as the zero-day share rises. In Mandiant's data the ratio of n-days to zero-days moved from 38:62 in 2021-2022 to 30:70 in 2023. A negative average does not mean patch analysis became faster than instant. Second, the 5-day figure for 2023 excludes 15 outliers and would be 47 days with them. No 2024 figure could be confirmed from a primary Mandiant source.
That is why figures for the period after a patch only show the speed of patch diffing more directly. Mandiant's 41 n-days from 2023 break down as follows.
| Time since patch release | Share of n-days exploited (cumulative) |
|---|---|
| Within 1 day | 12% (5) |
| Within 1 week | 29% (12) |
| Within 1 month | 56% |
| Within 6 months | 95% (all but two) |
In the same analysis, where a public exploit appeared before exploitation in the wild, the median time from disclosure to the first public exploit was 7 days. Within a week, in other words, someone has read the patch and posted working code. The analysis also recorded a fact that cuts the other way. Of 2023 vulnerabilities with at least one public proof of concept or exploit, 72% were not known to be exploited in the wild. An exploit existing and an exploit being used are different things.
Other organizations' numbers point the same direction. They measure different things and should not be compared with each other directly.
| Source | What is measured | Figure |
|---|---|---|
| Rapid7 2026 Global Threat Landscape Report (March 18, 2026) | Exploited high- and critical-severity vulnerabilities | 71 (2024) → 146 (2025), up 105% |
| Same report | Time from publication to CISA KEV listing, high and critical flaws | Median 8.5 → 5.0 days; mean 61.0 → 28.5 days |
| VulnCheck first-half 2026 report (July 28, 2026) | Share with exploitation evidence on or before the CVE publication date | 28.93% in 2025 → 23.43% in the first half of 2026 |
| VulnCheck first-half 2025 report (July 30, 2025) | Same measure | 23.6% in 2024 → 32.1% in the first half of 2025 |
| VulnCheck first-half 2026 report | Median time from CVE publication to exploited-list entry | 120 days (2025) → 80 days (first half of 2026) |
| Same report | CVEs reaching the exploited list within 31 days of publication | 196 (2024), 194 (2025), about 200 (first half of 2026) |
Rapid7's chief scientist Raj Samani said in the release that exploitation timelines are increasingly measured in days rather than weeks. The last VulnCheck row deserves attention. About 190 vulnerabilities a year reached its exploited list within a month in 2024 and 2025. In 2026 about 200 did so in six months. The three organizations define things differently, and their conclusions converge: the period in which it is safe to wait after a patch is counted in days, not months.
3. Exploit Wednesday: a 20-year-old problem
That a patch hands over a clue is not a new discovery. Since October 2003 Microsoft has shipped its security updates together on the second Tuesday of each month, a schedule known as Patch Tuesday. According to Wikipedia, batching was meant to reduce distribution costs and give administrators time to prepare. A fixed date also meant analysts could wait for it. Patches arrived on Tuesday, exploits derived from them appeared soon after, and the day that followed picked up the name Exploit Wednesday. Wikipedia cites an October 2006 Trend Micro blog post and an October 2007 Redmond Magazine article for the term.
A glossary page from the software company ManageEngine adds that the term does not mean every flaw is exploited the next day, or at all. Its point is that once patches are public, researchers and attackers can compare vulnerable and patched software to see what the fixes address, and endpoints stay exposed for as long as they remain unpatched. In a month like the November 2026 Patch Tuesday covered in our November 11, 2026 report, when more than 1,000 fixes ship at once, defenders are still deciding what to install first while an analyst only needs to pick the few fixes worth the effort.
That the process could be automated was also shown long ago. In 2008 four researchers including David Brumley of Carnegie Mellon University and Dawn Song of UC Berkeley published a paper on automatic patch-based exploit generation (APEG). According to The Register's coverage at the time, working from five Microsoft patches, their system took from six seconds to about three minutes to generate inputs that reach the vulnerable point once the differences between patched and unpatched binaries had been found. The caveats were that it did not produce fully weaponized exploits and might not work for every type of vulnerability. The researchers' conclusion was a warning about how patches are distributed: when an update rolls out over days, those who receive it first get time to attack those who have not. Among their proposals was to distribute patches in encrypted form ahead of time and release the key to everyone at once.
| Year | What happened | The human's role |
|---|---|---|
| October 2003 | Microsoft starts Patch Tuesday | Analysts reverse-engineer patches by hand |
| 2006-2007 | The term Exploit Wednesday appears in industry writing | A skilled analyst, one to several days |
| April 2008 | APEG paper: inputs generated in six seconds to about three minutes for five patches | People prepare the diff and choose targets |
| April 2024 | University of Illinois researchers: a GPT-4 agent exploits 87% of 15 one-day vulnerabilities | A person hands over the CVE description |
| August-September 2026 | PaperCut flaws: researchers and Huntress generate working exploits with LLMs within a day of disclosure | People direct, the LLM writes |
| November 2026 | GTIG reports exploit_agent | No human involvement after setup |
4. What AI shortened: what exploit_agent changes and what it does not
In April 2024 Richard Fang, Daniel Kang and colleagues at the University of Illinois Urbana-Champaign tested LLM agents on 15 real one-day vulnerabilities. According to the paper's abstract, GPT-4 exploited 87% of them when given the CVE description, while every other model tested at the time and the open-source vulnerability scanners scored 0%. Without the CVE description, GPT-4's success rate fell to 7%. That gap is the heart of the matter. A model from two years ago could do most of the job when told where the problem was, and could do almost none of it when it had to find the problem alone. A patch and its advisory are exactly the material that says where the problem is.
GTIG's November 2026 report makes the same argument as attacker economics. Threat actors may find it more efficient to use LLMs to rapidly weaponize patched n-days, it wrote, because they can focus research on a finite set of source materials: comparisons of source code versions, release notes, security research and proof-of-concept code for specific flaws. Hunting for zero-days requires casting a much wider net. Statistics GTIG analyzed in September 2026 fit this picture. Overall disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits rose only marginally, from a 2025 average of 8 a month to 11 a month in 2026, before spiking to 22 in August.
exploit_agent sits at the end of this line. What GTIG disclosed:
| Item | What GTIG said |
|---|---|
| How it was found | On an exposed staging server |
| Architecture | A multi-agent framework integrating several frontier and local models |
| What it was set up to do | Continuous differential analysis of popular web application software: comparing code before and after patch announcements, identifying potential vulnerabilities, drafting exploits and executing them automatically |
| Human involvement | None. GTIG wrote that this was the first time it had seen discovery and exploitation fully automated without a human in the loop |
| Also on the server | Data taken from victim organizations, and scripts for internal reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells |
| Targeting | Signs of interest in French websites in media and entertainment and in the education sector |
| Operator clues | Russian-language comments in a configuration file, telling the operator to enter LLM API keys in the file and not in the chat |
A figure in the report shows the framework examining access-permission checks in a WordPress plugin, which indicates the kind of software it went after. Not operating systems or browsers with deep defenses, but web software that is widely installed, has public source code and depends on each site operator to update it.
What AI shortened becomes clear from the table in section 1. Step 2, extracting differences, was already done by tools. Human time went into step 3, filtering noise, step 4, understanding the cause, and the code-writing that follows. That is the stretch LLMs now cover. The PaperCut case of late August 2026 shows the result. According to GTIG, two vulnerabilities in the print-management software PaperCut (CVE-2026-81578 and CVE-2026-82078) were first exploited as zero-days. Once they were disclosed, researchers and the security firm Huntress used LLMs to generate working exploits within a day, and four days after disclosure GreyNoise reported a campaign against hundreds of organizations in which the threat actor likely used LLMs to build and test exploits and assemble a target list.
Some things have not changed.
| Detail | Basis | |
|---|---|---|
| Changed | The human time needed to analyze one patch is gone, and there is no ceiling on how many software products can be watched at once | GTIG: configured for continuous differential analysis |
| Changed | The barrier to entry. Running a framework does not require a reverse-engineering specialist | GTIG cites a low barrier to entry and scaling potential |
| Unchanged | It does not work against systems that have applied the patch. That is the definition of an n-day | Mandiant's definition |
| Unchanged | Autonomous execution leaves many traces. GTIG wrote that these tools currently trade stealth for velocity | GTIG |
| New burden | A wave of LLM-generated exploits that are claimed to work but do not, adding verification work for defenders | GTIG, on the October 2026 PHP flaws |
The case behind that last row is the pair of PHP vulnerabilities (CVE-2026-109618 and CVE-2026-109619) covered in our October 26, 2026 report on PHP2Shell. According to GTIG, a proof-of-concept video and early patched code were published before the official patches of October 21, 2026, which let both researchers and attackers compare code versions and understand the flaws. Diffing began before the patch was formally out. In the days after the official patch, hundreds of non-functional proofs of concept were posted, and the security community spent effort sorting real from fake.
GTIG assessed it as highly likely that threat actors will keep using such frameworks in campaigns against organizations worldwide. It expects actors focused on long-term stealth, such as espionage groups, to rely less on forensically noisy tools, and actors less concerned with stealth, including many cybercrime groups, to adopt them first.
5. Defenders use the same technique
Patch diffing began as a defender's tool, and defenders use it far more often than attackers do.
| Use | What is done | Why it matters |
|---|---|---|
| Patch verification | Check that the flaw the vendor says it fixed is fixed, and that the same mistake is not left in neighboring code | Incomplete patches lead to bypass vulnerabilities |
| Writing detections | Learn the location and conditions of the flaw and build intrusion detection and blocking rules | Protects systems that cannot be patched right away |
| Finding silent fixes | Identify security flaws fixed without an advisory or a CVE number | Analysts can see fixes that are not announced; this keeps defenders from being the only ones who do not know |
| Prioritizing | Gauge real exploitation difficulty that a one-line advisory cannot convey | Decides what to install first out of hundreds of fixes |
| Scoping impact | Check whether other products that reuse the code, and older unsupported versions, share the flaw | A version that gets no patch keeps the flaw for good |
The last row connects to our October 26, 2026 explainer on end of life. When a supported version gets a patch, that patch also reveals that unsupported older versions have the same flaw. Wikipedia's Patch Tuesday entry notes that newer patches can reveal flaws in older Windows versions that are no longer supported. Since no patch will come for those versions, such systems become n-day targets with no end date.
AI is used on both sides too. The GTIG report says Google uses Big Sleep, an agent built by Google DeepMind and Project Zero, to find unknown vulnerabilities, and agents such as CodeMender to find and fix vulnerabilities before and after release. In a post published the same day, Google wrote that defending against machine-speed threats at human speed is not feasible. That sentence appears in a post promoting Google's own security products, which should be kept in mind, but the direction matches the revision CISA made to its own directive.
6. What individuals and organizations should do: automatic updates, then KEV and EPSS
If patch diffing cannot be stopped, one response is left: apply the patch before the analysis is finished. For individuals that is simple. For organizations it is a question of ordering.
| Who | What to do | Why |
|---|---|---|
| Individuals | Turn on automatic updates for the operating system, browser and phone | The time an update sits uninstalled is the n-day exposure window |
| Individuals | Restart the browser after it updates | A downloaded update that has not been applied leaves the old code running |
| Individuals | Replace devices and programs that are out of support, or take them off the internet | A flaw that gets no patch stays for good |
| Individuals | Check firmware auto-update on routers, NAS boxes and similar devices | Devices that are installed once and forgotten stay vulnerable longest |
| Site operators | Turn on automatic updates for the CMS core and plugins, WordPress included, and delete unused plugins | Popular web application software is what exploit_agent was aimed at |
| Companies | Keep an inventory of internet-facing assets and set patch deadlines for them in days | Exposed assets are what automated attacks reach first |
| Companies | Order installation by KEV listing and EPSS score | A CVSS score alone does not say what is being attacked |
| Companies | If a patch cannot go in now, isolate, block or disable the feature first | CISA also counts isolation and mitigation as remediation |
| Companies | Verify a proof of concept from the internet before it triggers an emergency change | Much LLM-generated code does not work |
The two indicators on the company side need explanation. As our October 6, 2026 explainer on CVSS set out, CVSS measures how severe a flaw is and says nothing about how likely it is to be attacked. KEV and EPSS fill that gap.
| Indicator | Run by | What it says | How to read it |
|---|---|---|---|
| CVSS | FIRST | Severity of the flaw, 0 to 10 | A high score does not mean it will be attacked |
| KEV | CISA, the US cybersecurity agency | A list of vulnerabilities confirmed to have been used in real attacks | If it is listed, it goes first. The catalog had 1,781 entries when checked on November 18, 2026 |
| EPSS | The EPSS special interest group at FIRST | Probability of exploitation in the next 30 days, 0 to 1, updated daily | Higher probability first. There is no official threshold |
KEV records the past. Only vulnerabilities with evidence of use in attacks are listed. The Outlook flaw from the November 2026 Patch Tuesday, CVE-2026-124194, and Chrome's CVE-2026-125350 were both added on November 10, 2026. EPSS forecasts the future. FIRST describes it as a model that estimates the probability that a published CVE will be exploited in the wild in the next 30 days. Scores are generated by a company called Empirical Security and published free every day. FIRST's guidance says scores cluster near zero, so a probability of 0.10 sits around the 95th percentile, and that 0.10 is commonly cited as a cut-off but carries no special authority from EPSS. Each organization should choose a threshold to fit its remediation capacity. When a CVE is on KEV and has a low EPSS score, the guidance is to follow KEV. EPSS speaks only to the likelihood of a threat. It knows nothing about how important or how exposed your own assets are, and that judgment stays with the organization.
The US government has shortened its deadlines. On June 10, 2026, CISA issued Binding Operational Directive BOD 26-04, revoking its 2019 and 2021 directives. According to Rapid7's summary, the 2021 directive gave a flat 14 days for every KEV entry. The new one sets five tiers based on four factors: whether the asset is publicly exposed, whether the flaw is on KEV, whether exploitation can be automated, and whether it gives total control of the system.
| Tier | Conditions (as summarized by Rapid7) | Deadline |
|---|---|---|
| 1 | On KEV and grants total control | 3 days, plus forensic triage for compromise |
| 2 | Publicly exposed, automatable, grants total control, not yet on KEV | 3 days |
| 3 | On KEV with partial control, or publicly exposed with partial control | 14 days |
| 4 | Internal asset, automatable, partial control | 60 days |
| 5 | Internal asset, not on KEV, not automatable, partial control | At the next system upgrade |
Deadlines run in calendar days from whichever comes first: the day CISA adds the CVE to KEV or the day the agency identifies the flaw on an asset. The forensic triage attached to tier 1 is where the directive meets this article's subject. It assumes that even if a system is patched within three days, it may already have been breached during those three days. According to CyberScoop, CISA officials cited AI assisting both researchers and adversaries in finding software flaws as background, and pointed to a Verizon report figure that only 26% of KEV-listed vulnerabilities were fully remediated in 2025, down from 38%. The directive binds US federal civilian agencies only. The three-day benchmark for internet-facing assets is one any organization can borrow.
7. Frequently asked questions
Is patch diffing illegal? The technique is a standard analysis method in the security industry, and the main tools, BinDiff and Diaphora, are publicly distributed. Defenders use it to verify patches and write detection rules. What is illegal is not the analysis but using its results to break into someone else's system without permission. Reverse engineering of commercial software can be restricted by national law and license terms.
Could vendors just hide what they fixed? No. To apply a patch the changed code has to reach the user, and received code can be compared. A vague advisory takes information away from defenders only, since analysts read the code regardless. When researchers in 2008 proposed distributing encrypted patches ahead of time and releasing the key simultaneously, the aim was not concealment. It was to have everyone receive the fix at the same moment.
If a patch gives attackers a clue, would it be better not to release it? No. Without a patch the flaw remains a zero-day and users have no defense at all. With a patch, at least everyone who updates is safe. Releasing a patch does not create the risk. It moves the risk onto those who have not updated.
Are automatic updates enough? For individuals they are the single most effective step. Two limits remain: many downloaded updates take effect only after a restart, and devices out of support receive nothing. Some products also roll automatic updates out in stages over several days, so when you hear of an urgent security update it is faster to check the settings screen yourself.
What if an update I install right away breaks something? That is why companies test patches before deploying them. The test period used to be allowed to run for weeks. On the statistics it is now a matter of days. A common compromise is to patch internet-facing systems first and internal systems after a short test. Where a patch cannot be applied yet, mitigations such as blocking or disabling the feature come first.
Are one-day and n-day different things? They name the same category. N-day covers any vulnerability after its patch is public, and one-day emphasizes the period right after disclosure.
Is open source more dangerous? The change history is public, which does make comparison easier, and GTIG wrote that attackers compare source code versions. But commercial products shipped only as executables are analyzed through binary diffing as well. The difference is the time the analysis takes, and tools and AI are shrinking that time. The conclusion is the same in both cases: apply the patch quickly once it is out.
Is AI also finding and attacking zero-days automatically? The exploit_agent framework in GTIG's November 2026 report is an n-day tool that compares code before and after patch announcements. The report says zero-day hunting requires a far wider search and counts only a marginal rise in zero-day exploits compared with the overall total. On how often AI-found vulnerabilities are exploited, VulnCheck's first-half 2026 report says that of 1,061 vulnerabilities it attributes to AI-assisted discovery, 14, or 1.3%, have been confirmed exploited in the wild.
How do I know whether my site is a target? Targets are picked by scanners, not by people. A vulnerable version of widely used software left open to the internet goes on the list whatever the size of the site. Checking the versions of your CMS and plugins and applying any security update that is out is the whole of the check.
8. What remains open or unconfirmed
| Item | Confirmed | Not confirmed |
|---|---|---|
| What exploit_agent achieved | Victim data was on the server, with signs of interest in French sites | The number of exploits built, the success rate, the time from patch to attack, the CVEs targeted |
| Report date | The Register's coverage and Google's summary post are dated November 16, 2026 | The Google Cloud Blog report was dated November 17 when checked on November 18, 2026 |
| Mandiant's −7 days | The summary sentence for M-Trends 2026 on the Google Cloud Blog | Sample size, outlier handling and the 2024 value; the full report was not opened |
| What the average means | That it includes zero-days, by definition | The 2024-2025 distribution for n-days alone |
| Rapid7 figures | Median 8.5 days to 5.0 days | The release does not state the years compared |
| VulnCheck figures | 28.93% for 2025, 23.43% for the first half of 2026 | Past values change as evidence is added; the first-half 2026 value may still rise |
| BOD 26-04 tier table | The three-day deadline, forensic triage, issue date and revocation of earlier directives | The detailed conditions of the five tiers follow Rapid7's summary; CISA's own table is an image |
| Origin of Exploit Wednesday | Wikipedia cites 2006 and 2007 articles | Who used it first; neither original could be opened |
| Size of the time AI saved | One day and four days in the PaperCut case | No measurement compares AI and human analysts under the same conditions |
| Detection indicators | GTIG says detection rules have shipped for another, publicly known framework | Whether indicators for exploit_agent itself have been published |
For why a vulnerability with no patch is the most dangerous kind, see our explainer on zero-days. How severity scores are assigned, and why they should not decide patch order alone, is in our explainer on CVSS. The problem of versions that no longer receive patches is in our explainer on end of life. Which of a month's 1,000-plus fixes to install first is worked through in our report on the November 2026 Patch Tuesday and CVE-2026-124194, and the mechanics of the flaw most often fixed in browser patches are in our explainer on use-after-free.
Sources
- Google Cloud Blog (GTIG) — GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold (2026-11)
- The Register — Google: AI agent turned a security patch into a working exploit with no human help (2026-11-16)
- Google blog — Outpace automated cyber threats with security built for nonstop defense (2026-11-16)
- AI Weekly — Google flags exploit_agent as first autonomous AI attack in wild (2026-11-16)
- Google Cloud Blog (Mandiant) — How Low Can You Go? An Analysis of 2023 Time-to-Exploit Trends (2024-10-16)
- Google Cloud Blog (Mandiant) — M-Trends 2026 (2026-03-24)
- Security Boulevard — Mean time to exploit has gone negative (2026-05-05)
- VulnCheck — State of Exploitation 1H 2026 (2026-07-28)
- VulnCheck — State of Exploitation 1H 2025 (2025-07-30)
- Rapid7 press release — 2026 Global Threat Landscape Report (2026-03-18)
- Rapid7 blog — Racing the three-day clock: CISA BOD 26-04 (2026-11-04)
- CISA — BOD 26-04: Prioritizing Security Updates Based on Risk (2026-06-10)
- CISA — Known Exploited Vulnerabilities Catalog
- CyberScoop — CISA vulnerability remediation directive BOD 26-04 (2026-06)
- FIRST — Exploit Prediction Scoring System (EPSS)
- FIRST — EPSS frequently asked questions
- GitHub google/bindiff — README
- GitHub joxeankoret/diaphora — README
- Wikipedia — Patch Tuesday
- ManageEngine — Exploit Wednesday glossary page
- The Register — MS patch system poses significant risk, say researchers (2008-04-25)
- arXiv 2404.08144 — LLM Agents can Autonomously Exploit One-day Vulnerabilities (2024-04)