WIRED’s “AI-Fueled Ransomware Boom” Rests on Leak-Site Counts That Reproduce at 25% to 39%, but the ReliaQuest Report Behind Its New-Group Numbers Calls the Rise Unlikely to Reflect a Real Increase in Activity and Never Mentions AI, and in a Public Tracker 97% of the Rise Comes From Groups First Seen in Q3
WIRED says an AI-fueled ransomware boom is under way. Four counts agree that Q3 2026 leak-site listings rose 25% to 39%, but ReliaQuest’s report calls the rise unlikely to reflect a real increase in...
WIRED’s Kernel Panic newsletter of 14 November, “Beware the New AI-Fueled Ransomware Boom,” says leak-site claims have spiked, that AI is helping newcomers start their own ransomware groups, and that fully AI-powered ransomware has not arrived yet. It quotes researchers at Halcyon, ReliaQuest and Recorded Future and links the Q3 2026 reports from Halcyon and ReliaQuest. I checked its numbers against those two reports, against Comparitech’s quarterly count, and against my own count of 23,208 listings from the public ransomware.live tracker, with data through 12:25 UTC on 14 November.
Table Of Content
- The claims and what I found
- What the newsletter claims, and what it hedges
- The counts reproduce, and the spread has a simple cause
- The rise has not reversed
- Who posted the rise
- “New” is a loose word
- Same groups, different counts
- What a listing proves
- Where AI enters the evidence
- What WIRED’s sources say
- What the two reports say
- What a count can and cannot show
- The United States and the “international” claim
- “Already dwarfs the total number for 2025”
- Reproduce it
- Run it
- What each section checks
- The script
- Its output on my snapshot
- If a leak site names your organization
- What this audit cannot show
- Sources
The counts hold up. The cause does not come from the counts. The ReliaQuest report that supplies WIRED’s numbers on new groups says the 30% rise “is unlikely to reflect a real increase in ransomware activity” and never mentions AI. Halcyon’s link between AI and the rise is an inference, not a measurement. And in my count, 97% of the Q3 rise comes from 34 groups the tracker had never recorded before July.
The claims and what I found
| WIRED’s claim | What the reports and the tracker show | Verdict |
|---|---|---|
| “Nearly 3,000 attacks over the summer,” up “20 to 40 percent” on spring | Halcyon 2,766 (+39.1%), Comparitech 2,627 (+29%), ReliaQuest +30%, tracker 2,958 (+25.3%) | Holds; most of the spread is a different Q2 baseline |
| “Around 111 ransomware groups” made claims | Halcyon’s report says 111; the tracker has 107 active groups in Q3 | Holds |
| Global Secret Group listed 48 victims, more than half on one day; Storm listed 71 in 20 days | ReliaQuest says so; the tracker has 48 (29 on one day) and 78 in Q3 | Close; crpx0, another ReliaQuest example, differs by 10 listings |
| The boom is AI-fueled | ReliaQuest: unlikely to be a real rise in activity, no mention of AI. Halcyon: “hard to imagine” it was not partly AI, an inference | Cause not measured |
| Most of the increase is international | US share of listings 48.9%, 37.8%, 36.6% across Q1 to Q3 in the tracker (ReliaQuest: 55%, 49%, 41%) | Holds in direction; levels differ by tracker |
| 2026 “already dwarfs” 2025 | Tracker: 9,349 listings by 14 November against 8,190 for all of 2025 (+14.2%) | Exceeded; “dwarfs” is generous; encryption is not visible in listings |
What the newsletter claims, and what it hedges
The piece is more careful than its headline. It concedes that “it is difficult to determine exact numbers of ransomware attacks, given that victims don’t always disclose incidents and attackers wildly inflate their claims,” says “fully AI-powered ransomware attacks have yet to emerge,” and adds that “not all of the attacks may be legitimate.” Within those limits it makes six claims that can be checked:
- Leak-site claims reached “nearly 3,000 attacks over the summer,” which some researchers read as “a 20 to 40 percent increase since the spring.”
- Halcyon’s analysis counts “around 111 ransomware groups” claiming attacks between July and September.
- New groups arrived fast. ReliaQuest’s analysis has Global Secret Group claiming “48 victims, with more than half of these being listed on the same day,” and Storm listing “71 victims in just 20 days.”
- The cause is partly AI. Cynthia Kaiser of Halcyon says “I do think it is in part fueled by AI,” and Brandon Tirado of ReliaQuest says the growth happened because of “this acceleration that AI has provided to these threat actors, essentially allowing for them to go out and start their own groups.” Allan Liska of Recorded Future says AI is used for “better phishing emails and in the negotiation process, but the ‘successful’ ransomware groups are still largely human operated.”
- Most of the increase is international: Kaiser says “the United States was a little more steady.”
- Liska says the number of attacks Recorded Future has seen this year “already dwarfs the total number for 2025,” while the number “involving encryption may actually decrease year over year.”
The counts reproduce, and the spread has a simple cause
Four counts of Q3 2026 listings are available. They agree on direction and differ by about 14 points on the size of the rise:
| Source | Q3 2026 | Q2 2026 baseline | Change | Active groups in Q3 |
|---|---|---|---|---|
| Halcyon | 2,766 claims | 1,988 (its Q2 report) | +39.1% | 111 |
| Comparitech | 2,627 attacks | 2,030 | +29% | not stated |
| ReliaQuest | not stated in the text | not stated | +30% | not stated (43 first-time groups) |
| ransomware.live, my count | 2,958 listings | 2,360 | +25.3% | 107 |
Most of that spread sits in the Q2 baseline, not in Q3. The trackers are 13% apart on Q3 (2,627 to 2,958) and 19% apart on Q2 (1,988 to 2,360); both percentages are my own division of the figures above. Halcyon’s Q2 was a dip, down 5.7% on Q1 by its Q2 report, while in the tracker Q2 was flat on Q1 (+0.7%), so a similar Q3 volume reads as a smaller rise there. WIRED’s “20 to 40 percent” brackets all four counts, and “nearly 3,000” fits both Halcyon (2,766) and the tracker (2,958). Infosecurity Magazine’s write-up of Comparitech’s count says 27%; I use the 29 percent on Comparitech’s own page.
The rise has not reversed
The weekly shape matters for any story about acceleration. Counting listings in 7-day windows, the 13 weeks from 1 April averaged 182 a week and the 13 weeks from 1 July averaged 223. Inside Q3, the five weeks from 1 July averaged 212, the four weeks from 5 August 273 and the four weeks from 2 September 186, close to Q2’s average. October and November have been higher than any of those: 1,688 listings in the 45 days to 14 November is about 262 a week. So the rise was a step with an August peak and a September dip, and the pace in Q4 so far is above Q3’s.
Who posted the rise
ReliaQuest’s central finding is about composition. “Groups appearing for the first time in Q3 accounted for 93% of this quarter’s uptick in named organizations,” it says, while groups active in both Q2 and Q3 rose “by only 1%,” and the 43 new groups “had never posted before Q3.” The tracker reproduces the pattern with different numbers. Splitting the Q2 to Q3 rise of 598 listings by what each group did:
| Kind of group | Groups | Change in listings | Share of the rise |
|---|---|---|---|
| Never recorded before Q3 | 34 | +580 | 97% |
| Back after a quiet Q2 | 5 | +61 | 10% |
| Active in both quarters | 68 | +72 | 12% |
| Posted in Q2, silent in Q3 | 21 | -115 | -19% |
The 34 new groups posted 580 of the 2,958 Q3 listings (19.6%). The established groups moved in both directions: The Gentlemen went from 252 to 395 listings and Qilin from 297 to 362, while Akira fell from 124 to 89 and 21 groups that posted in Q2 went quiet, so the net for groups active in both quarters is +72, a 3.2% rise on their 2,245 Q2 listings against ReliaQuest’s 1%.
The clock matters. I counted each listing on the date the tracker first saw it, the closest proxy for when a name appeared on a leak site. The tracker also files listings in monthly files keyed mostly by the attack date a group claims. Using those files as the quarter instead, the same four rows read 59%, 11%, 48% and -18% of a 606-listing rise, which suggests that groups listing older attacks push listings into earlier months. That nearly all of the growth comes from new groups therefore depends on counting by posting date, and neither report says which date it uses.
“New” is a loose word
ReliaQuest says the growth “was driven almost entirely by unverified new groups, several of which show signs of being rebrands, affiliate breakaways, or exaggerated claims rather than genuine new operations.” Its example is Storm, which “first posted about five weeks after” a similarly named group, Stormous, made its last post, a timing that makes a rebrand “a realistic possibility, though it’s unconfirmed.” The tracker shows a similar gap: its last Stormous listing is 28 June and its first Storm listing 7 August, 39 days apart. Rapid7’s Christiaan Beek, quoted by Cyber Daily, puts the general point this way: “a new ransomware name does not mean a new ransomware operation.”
Same groups, different counts
The four newcomers that ReliaQuest names can be set against Halcyon and the tracker. I matched names by sight; the tracker spells the fourth crpxo with a letter o where ReliaQuest writes crpx0.
| Group | ReliaQuest | Halcyon | ransomware.live, Q3 (busiest day) |
|---|---|---|---|
| Global Secret Group | 48 victims, 28 on one day | 43, nearly two-thirds in July | 48 (29 on 26 July) |
| Storm | 71 over 20 days | 39 in August plus 35 in September | 78 (11 on 30 September) |
| Orova | 47 over 10 days | not listed | 47 (24 on 4 August) |
| crpx0 | 47, 41 on one day | not listed | 37 (20 on 27 July) |
Three of the four land within about 10% of ReliaQuest’s figure. crpx0 does not: 37 listings and 20 on its busiest day against 47 and 41. Halcyon’s 43 for Global Secret Group is five below the other two. The lesson is that claim counts for the same group differ by up to about 20% depending on who is counting, and the busiest-day figure is more fragile still, because a tracker that first scrapes a new site records everything it finds that day (Orova’s 24 listings on 4 August may be such a first sweep).
ReliaQuest reads single-day dumps as “data likely compiled in advance, not real-time intrusions,” while noting that bulk releases “don’t prove claims are recycled or fabricated.” In the tracker, dumps are not only a new-group habit. On days when the same group listed at least 20 victims, new groups account for 17% of their Q3 listings (99 of 580) and groups active in both quarters for 14% (332 of 2,317); at a threshold of 10 victims it is 37% against 28%. DeadLock, which the tracker had already recorded in Q2 and so counts as established here, bursts too: Halcyon counts 87 claims in July and none in September, and the tracker has 91 listings for it in Q3 against 10 in Q2.
What a listing proves
Comparitech splits its Q3 total into “247 confirmed attacks (confirmed by the entity involved)” and 2,380 unconfirmed ones. By my arithmetic that is 9.4% confirmed. ReliaQuest’s clearest example of a thin claim is ExfilSquad, whose published data contained “low-risk workforce and customer relationship details that looked like they likely came from exposed credentials, not a network breach.” The tracker has 15 ExfilSquad listings, all first seen on 26 July.
The tracker can also flag two weaker signals, which are more common among new groups but remain a minority of their listings:
| Signal in Q3 listings | New groups (580) | Groups active in both quarters (2,317) |
|---|---|---|
| Same victim name already listed by another group since January 2024 | 35 (6.0%) | 52 (2.2%) |
| Claimed attack dated more than 90 days before the tracker saw the listing | 55 (9.5%) | 27 (1.2%) |
Name matching is crude, and a victim hit by two groups legitimately matches too, so these are base rates for context, not verdicts. They do fit ReliaQuest’s warning that listings “mix new attacks, rebrands, recycled data, and unsupported claims,” and Halcyon’s observation that “ten of the fifteen groups with the highest claim volume in August had not ranked there in July.”
Where AI enters the evidence
What WIRED’s sources say
The newsletter’s causal sentences are opinions from the people interviewed. Kaiser says the international growth is “in part fueled by AI.” Tirado goes further: the growth “has all happened because” of the acceleration AI gave “to these threat actors.” Liska, in the same piece, says the groups that succeed “are still largely human operated.” WIRED’s own sentence is the narrowest: “Fully AI-powered ransomware attacks have yet to emerge.” In my reading, the word “boom” and the “AI-fueled” label in the headline rest mainly on the first two statements.
What the two reports say
ReliaQuest’s report does not mention AI. I searched its body text, about 13,000 characters from the Key Points to the start of its product section, for AI, artificial, LLM, generative, agentic, automat and machine learning and found no match. Its explanation for the growth is “unverified new groups,” and its Key Points say that “victim counts are now an even less reliable measure of actual attacks.” Its section heading says it more bluntly: “Leak-Site Growth Doesn’t Equate to Attack Growth.”
Halcyon’s report is where AI appears. Its executive summary says “it is hard to imagine this increase was not at least in part due to expanded AI use among ransomware crews” and, in the next sentence, that “AI has not fundamentally changed how most of these groups attack.” The first is an inference: the report text I could read gives no measurement that ties AI use to the number of claims. The evidence it cites is case-level. One item is Anthropic’s September report on “operators suspected to be affiliates of the ShinyHunters collective,” in which a session-store dump of “over 2,100 Azure AD token sets spanning more than 40 corporate tenants” took “about 34 hours” and “AI agents performed nearly all of the work.” Anthropic describes that group’s activity as data theft followed by “pay-or-leak extortion demands.” The other item is Halcyon’s own October post: “of 405 malware samples connected to AI examined in one vendor study, only twelve had ever turned up on a real, protected computer.”
What a count can and cannot show
A leak-site count records that a name appeared on a page. It cannot say how the data was obtained, which tools were used, or whether the claim is true. The same Q3 numbers would appear if the new names were rebrands, affiliates moving between platforms, or pre-collected data. Beek says “the ransomware ecosystem is becoming more like a platform economy,” in which established operators supply infrastructure and affiliates can appear under their own brands. Halcyon’s observation about the August leaderboard fits churn at least as well as a steady stream of AI-enabled newcomers; that comparison is my reading, not Halcyon’s.
What would count as evidence is case-level: logs, transcripts and incident reports that show an agent doing the work, such as the Anthropic report above, the exposed Claude Code logs tied to an AI pentest agent at Korean financial firms, or WorkOS’s account of an agent that escaped a researcher’s sandbox, both covered here earlier. Those show AI inside intrusions. They do not show how many leak-site listings AI produced. Both vendors also sell products aimed at the threat they describe: WIRED introduces Kaiser as leading research for “the anti-ransomware platform Halcyon,” and ReliaQuest’s report closes with a pitch for its GreyMatter platform.
The United States and the “international” claim
Kaiser says “most of the recent increase was fueled by international attacks.” ReliaQuest backs it with figures: “The US share of all listed victims fell for the second consecutive quarter: from 55% in Q1, to 49% in Q2, to 41% in Q3,” with “approximately 85% of the added victims” outside the US. Halcyon puts the US at “more than 1,100 claims, about two in every five.” In the tracker, among listings that carry a country code, the US share is 48.9%, 37.8% and 36.6% for Q1 to Q3, and 68% of the Q2 to Q3 rise sits outside the US.
The direction matches in every source and the level does not: 41% for ReliaQuest, about 40% for Halcyon, 36.6% in the tracker. The tracker adds a wrinkle. Its big drop in the US share came in Q2, when total listings were flat, and Q3’s increase moved the share by only about one point. US listings grew 20.7% in Q3 against 24.7% for all listings with a country, which is steadier than the total but not by much. In the tracker, 218 of the 2,958 Q3 listings (7.4%) carry no country code.
“Already dwarfs the total number for 2025”
Liska’s sentence compares Recorded Future’s count for this year with its total for 2025. Recorded Future’s own figure for 2025 is that “publicly reported attacks rose to 7,200 in 2025 compared to 4,900 in 2024,” a count that “classifies both encryption attacks and data theft attacks with an extortion component under the ransomware umbrella,” with “approximately 50%” in the data-theft category (Recorded Future). I cannot see its 2026 count. The tracker has 8,190 listings for all of 2025. Its 2026 total passed that on 16 October and stands at 9,349 on 14 November: 14.2% above all of 2025 with about seven weeks of the year left, and 35.5% above the same span of 2025 (6,898). “Dwarfs” fits the same-span comparison better than the full-year one.
The second half of Liska’s sentence, that attacks involving encryption “may actually decrease,” cannot be tested from leak-site data. A listing does not say whether files were encrypted, and Recorded Future’s own taxonomy puts both kinds under one label.
Reproduce it
The script below uses only the Python standard library and was run on Python 3.13.14. It downloads the tracker’s monthly files from January 2024, deduplicates listings by group, victim and claim URL, and prints eight sections.
Run it
python leak_audit.py fetch data
python leak_audit.py report data
The tracker’s free API answers HTTP 429 now and then; the script waits 65 seconds and retries, so allow several minutes for the 35 monthly files. A re-run will not match my output exactly. The tracker adds late listings and revises old ones, and the Q4 lines change every day.
What each section checks
- Listings and active groups per quarter: WIRED’s “nearly 3,000,” the 20 to 40 percent range and the “111 groups.”
- The split of the Q2 to Q3 rise by group type, plus the same split on the tracker’s month files.
- Same-day dumps, and the named groups from ReliaQuest and WIRED.
- Repeated victim names and old attack dates.
- US share and the outside-the-US share of the rise.
- Weekly counts and the Q4 pace so far.
- Year to date against 2025, for Liska’s “dwarfs.”
- The Stormous to Storm gap.
The script
# leak_audit.py
"""leak_audit.py: check leak-site growth claims against ransomware.live monthly dumps (standard library only).
python leak_audit.py fetch data # download 2024-01 up to this month, waiting out HTTP 429
python leak_audit.py report data # print the tables used in the article
"""
import collections
import datetime as dt
import json
import re
import statistics
import sys
import time
import urllib.error
import urllib.request
from pathlib import Path
API = "https://api.ransomware.live/v2/victims/{y}/{m:02d}"
def fetch(folder):
folder = Path(folder)
folder.mkdir(parents=True, exist_ok=True)
today = dt.date.today()
y, m = 2024, 1
while (y, m) <= (today.year, today.month):
path = folder / f"{y}-{m:02d}.json"
for _ in range(8):
if path.exists():
break
try:
with urllib.request.urlopen(API.format(y=y, m=m), timeout=90) as r:
path.write_bytes(r.read())
print("saved", path.name)
except urllib.error.HTTPError as e:
if e.code != 429:
raise
print(path.name, "answered 429, waiting 65 s")
time.sleep(65)
if not path.exists():
sys.exit(f"gave up on {path.name}")
y, m = (y, m + 1) if m < 12 else (y + 1, 1)
def utc(text):
when = dt.datetime.fromisoformat(text.replace("Z", "+00:00"))
return when if when.tzinfo else when.replace(tzinfo=dt.timezone.utc)
def qtr(when):
return f"{when.year}Q{(when.month - 1) // 3 + 1}"
def load(folder):
seen, rows = set(), []
for path in sorted(Path(folder).glob("*.json")):
data = json.loads(path.read_text(encoding="utf-8"))
if not isinstance(data, list): # a saved rate-limit message
continue
for r in data:
key = (r.get("group"), r.get("victim"), r.get("claim_url"))
if key in seen or not r.get("discovered"):
continue
seen.add(key)
rows.append({
"group": (r.get("group") or "").lower(),
"when": utc(r["discovered"]),
"attack": utc(r["attackdate"]) if r.get("attackdate") else None,
"file_q": qtr(dt.datetime.strptime(path.stem, "%Y-%m")),
"country": (r.get("country") or "").upper(),
"name": re.sub(r"[^a-z0-9.]", "", (r.get("domain") or r.get("victim") or "").lower()),
})
rows.sort(key=lambda r: r["when"])
for r in rows:
r["seen_q"] = qtr(r["when"])
return rows
def pct(a, b):
return f"{100.0 * (a - b) / b:+.1f}%"
def partition(rows, field):
"""Split the Q2 to Q3 change by what each group did; field picks the quarter bucketing."""
first = {}
for r in sorted(rows, key=lambda r: r[field]):
first.setdefault(r["group"], r[field])
c2 = collections.Counter(r["group"] for r in rows if r[field] == "2026Q2")
c3 = collections.Counter(r["group"] for r in rows if r[field] == "2026Q3")
new = [g for g in c3 if first[g] == "2026Q3"]
back = [g for g in c3 if g not in c2 and g not in new]
both = [g for g in c3 if g in c2]
gone = [g for g in c2 if g not in c3]
rise = sum(c3.values()) - sum(c2.values())
parts = [
("groups never seen before Q3", new, sum(c3[g] for g in new)),
("groups back after a quiet Q2", back, sum(c3[g] for g in back)),
("groups active in both quarters", both, sum(c3[g] - c2[g] for g in both)),
("groups that posted in Q2 but not Q3", gone, -sum(c2[g] for g in gone)),
]
return rise, parts, c2, c3, new, both
def report(rows):
by_q = collections.Counter(r["seen_q"] for r in rows)
groups_q = collections.defaultdict(set)
for r in rows:
groups_q[r["seen_q"]].add(r["group"])
print(f"data: {len(rows)} listings, first seen {rows[0]['when']:%Y-%m-%d} to {rows[-1]['when']:%Y-%m-%d %H:%M} UTC")
print("\n1. Listings and active groups by quarter (date the tracker first saw the listing)")
for q in sorted(by_q):
if q >= "2025Q1":
print(f" {q}: {by_q[q]:>5} listings, {len(groups_q[q]):>3} groups")
q1, q2, q3 = by_q["2026Q1"], by_q["2026Q2"], by_q["2026Q3"]
print(f" Q3 vs Q2 {pct(q3, q2)} Q3 vs Q1 {pct(q3, q1)} Q2 vs Q1 {pct(q2, q1)}")
rise, parts, c2, c3, new, both = partition(rows, "seen_q")
print(f"\n2. Where the Q2 to Q3 rise of {rise} listings came from (lookback to {rows[0]['when']:%Y-%m-%d})")
for label, groups, delta in parts:
print(f" {label:<38} {len(groups):>3} groups {delta:>+5} listings = {100.0 * delta / rise:>4.0f}% of the rise")
posted = sum(c3[g] for g in new)
print(f" the {len(new)} new groups posted {posted} of the {q3} Q3 listings ({100.0 * posted / q3:.1f}%)")
base = sum(c2[g] for g in both)
print(f" groups active in both quarters rose {100.0 * sum(c3[g] - c2[g] for g in both) / base:.1f}% (on a base of {base} Q2 listings)")
print(" Q3 top groups:", ", ".join(f"{g} {n} (Q2 {c2[g]})" for g, n in c3.most_common(6)))
rise_f, parts_f, *_ = partition(rows, "file_q")
print(f" same split with the tracker's month files as the quarter (rise {rise_f}): " + ", ".join(f"{100.0 * d / rise_f:.0f}%" for _, _, d in parts_f))
day = collections.Counter((r["group"], r["when"].date()) for r in rows)
def bulk(groups, threshold):
mine = [r for r in rows if r["seen_q"] == "2026Q3" and r["group"] in groups]
big = [r for r in mine if day[(r["group"], r["when"].date())] >= threshold]
return len(big), len(mine)
print("\n3. Q3 listings on a day when the same group listed at least N victims")
for n in (10, 20):
a, b = bulk(set(new), n), bulk(set(both), n)
print(f" N={n:<2} new groups {a[0]}/{a[1]} = {100.0 * a[0] / a[1]:.0f}% groups active in both quarters {b[0]}/{b[1]} = {100.0 * b[0] / b[1]:.0f}%")
print(" named groups (Q3 listings, all listings, first seen, busiest day):")
for needle in ("global secret", "crpx", "storm", "orova", "exfilsquad"):
for g in sorted({r["group"] for r in rows if needle in r["group"]}):
mine = [r for r in rows if r["group"] == g]
days = collections.Counter(r["when"].date() for r in mine)
top = days.most_common(1)[0]
print(f" {g:<20} {sum(1 for r in mine if r['seen_q'] == '2026Q3'):>3} {len(mine):>3} {min(days)} {top[0]} = {top[1]}")
print("\n4. Q3 listings whose name was listed earlier, and listings that claim an old attack")
seen = collections.defaultdict(set)
kind = {}
for i, r in enumerate(rows):
others = seen[r["name"]] - {r["group"]} if r["name"] else set()
kind[i] = "other" if others else ("same" if r["group"] in seen[r["name"]] else "first")
if r["name"]:
seen[r["name"]].add(r["group"])
for label, groups in (("new groups", set(new)), ("groups active in both quarters", set(both))):
idx = [i for i, r in enumerate(rows) if r["seen_q"] == "2026Q3" and r["group"] in groups]
k = collections.Counter(kind[i] for i in idx)
dated = [rows[i] for i in idx if rows[i]["attack"]]
old = [r for r in dated if (r["when"] - r["attack"]).days > 90]
print(f" {label:<32} {len(idx):>4} listings: earlier by another group {k['other']} ({100.0 * k['other'] / len(idx):.1f}%), attack dated >90 days before first seen {len(old)} ({100.0 * len(old) / len(dated):.1f}%)")
print("\n5. Country (listings with a country code)")
net = {}
for q in ("2026Q1", "2026Q2", "2026Q3"):
known = [r for r in rows if r["seen_q"] == q and r["country"]]
us = sum(1 for r in known if r["country"] == "US")
net[q] = (len(known), us)
print(f" {q}: {len(known)} with a country, US {us} = {100.0 * us / len(known):.1f}%, no country {by_q[q] - len(known)}")
dn, du = net["2026Q3"][0] - net["2026Q2"][0], net["2026Q3"][1] - net["2026Q2"][1]
print(f" Q2 to Q3: US {pct(net['2026Q3'][1], net['2026Q2'][1])}, all {pct(net['2026Q3'][0], net['2026Q2'][0])}, outside the US = {100.0 * (dn - du) / dn:.0f}% of the rise")
print("\n6. Weekly shape (7-day windows from 1 April and from 1 July)")
for start in (dt.date(2026, 4, 1), dt.date(2026, 7, 1)):
weeks = [sum(1 for r in rows if start + dt.timedelta(7 * i) <= r["when"].date() < start + dt.timedelta(7 * i + 7)) for i in range(13)]
print(f" from {start}: {weeks} mean {statistics.mean(weeks):.0f}")
today = rows[-1]["when"]
days_q4 = (today.date() - dt.date(2026, 10, 1)).days + 1
q4 = [r for r in rows if r["seen_q"] == "2026Q4"]
first = {}
for r in rows:
first.setdefault(r["group"], r["seen_q"])
print(f" per day: Q2 {q2 / 91:.1f}, Q3 {q3 / 92:.1f}, Q4 so far {len(q4) / days_q4:.1f} ({len(q4)} listings in {days_q4} days)")
print(f" Q4 so far from groups first seen in Q3: {sum(1 for r in q4 if first[r['group']] == '2026Q3')}, in Q4: {sum(1 for r in q4 if first[r['group']] == '2026Q4')}")
span = lambda y: sum(1 for r in rows if r["when"].year == y and (r["when"].month, r["when"].day) <= (today.month, today.day))
total_2025 = sum(1 for r in rows if r["when"].year == 2025)
print(f"\n7. Same calendar span 1 Jan to {today:%d %b}: 2025 {span(2025)} listings, 2026 {span(2026)} ({pct(span(2026), span(2025))}); all of 2025: {total_2025} ({pct(span(2026), total_2025)} for 2026 so far)")
count = 0
for r in rows:
count += r["when"].year == 2026
if count > total_2025:
print(f" 2026 passed the 2025 total on {r['when']:%d %b}")
break
stormous = [r["when"] for r in rows if r["group"] == "stormous"]
storm = [r["when"] for r in rows if r["group"] == "storm"]
print(f"\n8. Last Stormous listing {max(stormous):%Y-%m-%d}, first Storm listing {min(storm):%Y-%m-%d}, {(min(storm) - max(stormous)).days} days apart")
if __name__ == "__main__":
if len(sys.argv) != 3 or sys.argv[1] not in ("fetch", "report"):
sys.exit(__doc__)
if sys.argv[1] == "fetch":
fetch(sys.argv[2])
else:
report(load(sys.argv[2]))
Its output on my snapshot
This is the output on the data I downloaded on 14 November, 23,208 listings with the latest first seen at 12:25 UTC.
data: 23208 listings, first seen 2024-01-01 to 2026-11-14 12:25 UTC
1. Listings and active groups by quarter (date the tracker first saw the listing)
2025Q1: 2372 listings, 68 groups
2025Q2: 1654 listings, 73 groups
2025Q3: 1693 listings, 77 groups
2025Q4: 2471 listings, 70 groups
2026Q1: 2343 listings, 72 groups
2026Q2: 2360 listings, 89 groups
2026Q3: 2958 listings, 107 groups
2026Q4: 1688 listings, 91 groups
Q3 vs Q2 +25.3% Q3 vs Q1 +26.2% Q2 vs Q1 +0.7%
2. Where the Q2 to Q3 rise of 598 listings came from (lookback to 2024-01-01)
groups never seen before Q3 34 groups +580 listings = 97% of the rise
groups back after a quiet Q2 5 groups +61 listings = 10% of the rise
groups active in both quarters 68 groups +72 listings = 12% of the rise
groups that posted in Q2 but not Q3 21 groups -115 listings = -19% of the rise
the 34 new groups posted 580 of the 2958 Q3 listings (19.6%)
groups active in both quarters rose 3.2% (on a base of 2245 Q2 listings)
Q3 top groups: thegentlemen 395 (Q2 252), qilin 362 (Q2 297), incransom 111 (Q2 100), krybit 95 (Q2 58), deadlock 91 (Q2 10), akira 89 (Q2 124)
same split with the tracker's month files as the quarter (rise 606): 59%, 11%, 48%, -18%
3. Q3 listings on a day when the same group listed at least N victims
N=10 new groups 216/580 = 37% groups active in both quarters 660/2317 = 28%
N=20 new groups 99/580 = 17% groups active in both quarters 332/2317 = 14%
named groups (Q3 listings, all listings, first seen, busiest day):
global secret group 48 57 2026-07-26 2026-07-26 = 29
crpxo 37 37 2026-07-09 2026-07-27 = 20
storm 78 89 2026-08-07 2026-09-30 = 11
stormous 0 109 2024-01-14 2025-10-27 = 7
orova 47 65 2026-08-04 2026-08-04 = 24
exfilsquad 15 15 2026-07-26 2026-07-26 = 15
4. Q3 listings whose name was listed earlier, and listings that claim an old attack
new groups 580 listings: earlier by another group 35 (6.0%), attack dated >90 days before first seen 55 (9.5%)
groups active in both quarters 2317 listings: earlier by another group 52 (2.2%), attack dated >90 days before first seen 27 (1.2%)
5. Country (listings with a country code)
2026Q1: 2229 with a country, US 1089 = 48.9%, no country 114
2026Q2: 2198 with a country, US 830 = 37.8%, no country 162
2026Q3: 2740 with a country, US 1002 = 36.6%, no country 218
Q2 to Q3: US +20.7%, all +24.7%, outside the US = 68% of the rise
6. Weekly shape (7-day windows from 1 April and from 1 July)
from 2026-04-01: [158, 246, 159, 199, 253, 228, 151, 125, 169, 144, 234, 163, 131] mean 182
from 2026-07-01: [213, 212, 164, 263, 207, 291, 255, 276, 269, 186, 183, 186, 189] mean 223
per day: Q2 25.9, Q3 32.2, Q4 so far 37.5 (1688 listings in 45 days)
Q4 so far from groups first seen in Q3: 419, in Q4: 176
7. Same calendar span 1 Jan to 14 Nov: 2025 6898 listings, 2026 9349 (+35.5%); all of 2025: 8190 (+14.2% for 2026 so far)
2026 passed the 2025 total on 16 Oct
8. Last Stormous listing 2026-06-28, first Storm listing 2026-08-07, 39 days apart
If a leak site names your organization
ReliaQuest’s advice is to treat a post from a new name as an unverified claim and let your own evidence decide. It asks three questions: “Was the named organization actually compromised?”, “Does the evidence support the actor’s account of how the data was obtained?” and “Does the name represent a distinct operation, a rebrand, an affiliate, or an opportunistic poster?” It warns that both errors cost something: an organization that acts on a false claim “may make an unnecessary payment, fund future criminal activity, and waste investigative resources,” while dismissing a real one leaves exposed data unaddressed.
The audit suggests four cheap signals to read from the tracker’s record for the group that named you: how old the group’s name is, how many victims it listed on the same day, whether your name was already listed by another group, and whether the claimed attack date is months older than the listing. None proves a claim false, and the base rates above show each is a minority pattern. Comparitech’s tally is the reminder that most listings are not confirmed by the organization named, so the check against your own logs is the step that settles it.
What this audit cannot show
- It uses one tracker. Its coverage differs from the vendors’ (107 Q3 groups against Halcyon’s 111, 34 never-seen-before groups against ReliaQuest’s 43, 37 crpx0 listings against 47), and its first-seen date for a group is when the tracker started recording it, which can be later than the group’s first post.
- The lookback starts in January 2024, so a name that last posted before then counts as new.
- Name matching uses the domain or victim name, so a victim legitimately hit by two groups counts as a repeat.
- The busiest-day figure depends on scrape timing, as the Orova row shows.
- Neither vendor publishes group-level data, so I could not reproduce ReliaQuest’s 43 or Halcyon’s 111 by name.
- AI use cannot be tested from listings at all.
- I did not check any claim about any victim and did not visit any leak site; the script reads only the tracker’s public API.
Sources
- WIRED, Kernel Panic, by Matt Burgess and Lily Hay Newman, 14 November 2026
- ReliaQuest, Ransomware and Cyber Extortion in Q3 2026, 3 November 2026
- Halcyon, Q3 2026 Ransomware Evolution Report and its Q2 2026 report
- Comparitech, Ransomware roundup: Q3 2026
- Anthropic, Countering misuse of AI: September 2026
- Halcyon, The AI Malware Panic Has a Measurement Problem
- Recorded Future, New ransomware tactics to watch out for in 2026
- Cyber Daily, Explainer: What’s driving the rapid rise in new ransomware groups?
- ransomware.live, public API, monthly victim files from January 2024 to 14 November 2026







No Comment! Be the first one.