Lead · Threat actors

Binance did not lose billions to a sophisticated attack. It lost them to a job interview

Binance says a fake recruitment process, a tampered Homebrew installer and an abused internal AI tool ended at its wallet signing service; the reported figure comes in three sizes and North Korea denies the whole thing.

confidence 0.85

Binance disclosed on 3 November that a theft it discovered in October was the end of what its chief executive called a months-long campaign: North Korean attackers who, in the company's own account, "combined a range of techniques, including social engineering by fake recruitment interviews", compromised engineers via a trojanized Homebrew installation, abused an internal AI tool, and reached the wallet signing service. North Korea denies responsibility. The number depends on which outlet you read on which hour: BleepingComputer reports about $6.34 billion, The Hacker News reports $6.43 billion, and the headlines round it to $6.3 billion.

Look at the chain and tell me where the hood is. A person who wanted a better job. An installer that looked like the one everybody on the team already runs. An internal tool, trusted because it is internal, doing what internal tools do, which is reach things. Nothing in that sequence required a zero-day, a submarine or a dark room with cascading green text. It required patience and a plausible recruiter, and the attackers had both.

The uncomfortable bit for the rest of us is the ordering: discovered in October, disclosed on 3 November, and still described in three different amounts on the day of disclosure. That gap is not a scandal, it is the normal physics of a large incident, and it is exactly the window in which your own board will ask you for a precise figure you do not yet have. Give them the chain instead. The chain is true earlier than the number is.

Why it mattersYour engineers' personal laptops, package managers and internal AI tooling are now part of the signing path for whatever you consider precious, whether or not that appears on the architecture diagram.

Today’s revelation

A supply chain is a list of strangers whose Monday becomes your Friday, and it is always Friday.

— Sturgeon’s Revelations · Probably Fine Daily No. 38

§2

Top stories

The rest of the last 24 hours, with receipts.

Vulnerabilities

New KEV entry: WordPress Core can be taken during the upgrade, which is the one moment you felt virtuous

CISA added CVE-2026-119977 to the Known Exploited Vulnerabilities catalogue on 3 November: a time-of-check time-of-use race condition in WordPress Core that could let unauthenticated attackers execute code on the server during the upgrade window, because of predictable, web-accessible staging. It carries an EPSS score of 0.006, which tells you how little the statistical models think of it and how much CISA's evidence of exploitation outranks them. One vulnerability added to the catalogue, that day, in that feed.

Why it mattersThe exploitable moment is the act of patching, so your maintenance window is now a detection window: watch for unexpected files under the staging path and for requests that arrive mid-upgrade.

confidence 0.90

Vulnerabilities

Node-RED is being exploited and it is wearing an industrial badge

Exploitation of CVE-2026-58115, an unauthenticated Node-RED admin API flaw, was detected from 30 October, with attackers deploying malicious flows that automatically download payloads; SecurityWeek reports multiple attacks from the same IP address dropping a Linux bot and webshells. The detail that should move this up your list is that Node-RED is embedded inside ICS products, so the thing that is exposed is frequently not something you bought as Node-RED. OTX carried it 2.3 hours before the news feeds did.

Why it mattersHunt for Node-RED admin endpoints you never knowingly deployed, particularly inside vendor appliances, and treat a flow you did not author as an incident rather than a config drift.

confidence 0.80

Vulnerabilities

A 9.8 in the Linux TCP stack is patched on one distribution and in a testing channel on another

Red Hat has shipped fixed kernels for a use-after-free in the Linux kernel's IPv4 TCP code, rated CVSS 9.8, for RHEL 8, 9 and 10, while Ubuntu's fixes for its three newest long-term-support releases sit in an opt-in testing channel with no date for a regular update; disabling MPTCP is explicitly not enough, because the bug is in IPv4 TCP. Separately, Barracuda disclosed a pre-authentication authorization bypass and a command injection that chain to root, with one of them affecting six products, both requiring access to the admin interface. And mSecure has told password-manager customers to consider changing passwords, resetting 2FA and removing passkeys after suspicious activity, without saying whether vaults were accessed.

Why it mattersPatch availability is not a single event: track per-distribution state for the kernel flaw, and treat 'requires admin interface access' as a reason to check who can reach the admin interface, not as a mitigation.

confidence 0.85

Ransomware

The leak sites had a very loud Tuesday, and most of the noise is one group posting in bulk

Ninja Signal's tempo count for today records 231 leak-site claims in seven days from 39 of the 271 groups tracked, with thegentlemen surging at 50 claims against 24.97 expected from its own prior rate, Booba Project at 11 against 2.86, and genesis at 9. DarkFeed's weekly roll-up puts the United States at 106 victims, Germany 19, and healthcare top by sector at 28, while Hackmanac counted 389 attacks across 52 countries for the week, 24 of them claimed by Qilin.

Why it mattersIf you report leak-site counts upward, say which are distinct victims and which are one actor's batch upload, or your trend line becomes a measure of a criminal's publishing schedule.

confidence 0.70

Threat actors

The update channel is the perimeter now: cPGuard shipped a backdoor to its own customers

An attacker broke into the update system of the cPGuard web security suite on 2 November and used it to push a malicious script to customers' Linux servers, installing a backdoor and collecting password hashes, database secrets and SSH files; the attacker retrieved files tied to 32 customer servers, and the affected versions are reported as 5.88.12 to 5.88.19. It is the same week Google dismantled the X.CPE botnet of more than 1.3 million Android devices, mostly TV set-top boxes, which it says device maker Videostrong intentionally backdoored and sold access to, and the same week Frontline Education disclosed a breach of over 1.3 million people through a flaw in third-party software, Seven-Eleven Japan confirmed 169,151 customer records exposed via logistics contractor CBcloud's PickGo platform, and Polymarket refunded about $2.7 million to 726 wallets after malicious code replaced a page-speed script it loaded from an outside company.

Why it mattersYour inventory of who can write to your servers without a change ticket is shorter than your vendor list and far more important than it: enumerate it today.

confidence 0.80

AI & security

LLMjacking is not a prank, it is a billing event with a business model behind it

Google warned on 3 November that attackers are hijacking enterprise access to large language models using stolen cloud credentials, running up fraudulent charges of up to $1 million at some organisations and more than $100,000 per day at others, then reselling the access through underground marketplaces; AlienVault's write-up places victims across North America, South America and Asia Pacific, including financial services. In the same twenty-four hours, WorkOS said an attacker used ARTEX AI to breach an internal sales app containing business contact records and summaries of customer communications, and Chainstack disclosed a breach of production systems and a customer database of names, email addresses and password hashes, which it assesses as part of a large-scale AI-powered campaign.

Why it mattersModel-inference spend is now a theft-detection signal: alert on cost and token-rate anomalies per key, and rotate cloud credentials that can call inference endpoints as if they were payment credentials, because they are.

confidence 0.85

AI & security

A 3AM affiliate appears to have had a model writing and revising its commands

A 3AM ransomware affiliate hit a UK school estate and encrypted around 37 of nearly 50 machines reachable from an unmanaged host, and researchers report evidence that a model may have generated and revised the commands during the intrusion, while stopping short of claiming full automation. The entry point is the part that will survive this news cycle: one host nobody owned, with line of sight to nearly fifty others.

Why it mattersHunt for the unmanaged host first; whatever wrote the commands, it still needed something on your network that no team has ever claimed.

confidence 0.75

Oddities

Second wave of supermarket fridge attacks across Europe, and a Russian emergency-alert app sending alerts it was not asked for

Overnight into Monday 2 November attackers again changed settings on internet-exposed refrigeration controllers at supermarkets in Belgium, Sweden, Estonia, Ireland and beyond, warming fridges and freezers; over 100 stores were affected, no chain reported more than several dozen, and some had to pull stock. It is the second such wave after an earlier one in France, and the perpetrators are unidentified. The same night, Russia's Ministry of Emergency Situations reported that an attacker exploited a vulnerability in the administrative panel of its official mobile application and pushed false alerts across 33 regions.

Why it mattersTemperature-controlled infrastructure, alerting apps and anything else where the payload is a setpoint rather than a file belongs in your exposure review, because the attack leaves almost no artefacts on anything you currently log.

confidence 0.80
§3

Hidden

The daily deep read: what is under the feeds, not on top of them.

Under the feeds: one actor's batch upload wearing forty-six hats, a CVE firehose that stopped mid-afternoon on Sunday, and vendor version numbers that do not agree with each other.

  1. Exception 01

    The thegentlemen spike is mostly tracker and story-generation artefact

    thegentlemen produced 46 items on 2026-11-03 against a median of about 0 on other days, N0n 22, Booba Project 13, Qilin 12 and Akira 12, with a 60-item cluster in the 12:00 hour against a mean of 7.6 per hour; many of the thegentlemen entries are same-timestamp Atlas 'cooling' stories rather than 46 distinct victims.

    confidence 0.55Speculative
    Wrong if
    Distinct named victims with distinct dates behind the thegentlemen items on the leak site itself.
    Watch
    Whether the thegentlemen count returns to roughly zero on 11-04 and whether Atlas keeps emitting bulk stories.
  2. Exception 02

    NVD went quiet after 2026-11-02T14:43

    All 100 NVD items in the window fall before 2026-11-02T14:43 and none after, which is most likely a feed or batch-ingest cutoff rather than a pause in CVE publication; treat CVE counts after that timestamp as unobserved rather than low.

    confidence 0.50Speculative
    Wrong if
    NVD items dated after 2 November 14:43 appearing on the next re-ingest.
    Watch
    The next NVD pull.
  3. Exception 03

    cPGuard's affected versions are reported inconsistently

    The hijacked-update story appears in three sources on the same day, with affected versions reported as 5.88.12 to 5.88.19 in one account and differently elsewhere; this matters because admins are being asked to self-assess exposure against a range that has not settled.

    confidence 0.60
    Wrong if
    A vendor advisory fixing a single authoritative affected-version range.
    Watch
    An OPSSHIELD advisory that settles the versions.
  4. Exception 04

    Mercado Bitcoin's no-leak statement sits next to a Breachforums advert

    Mercado Bitcoin reported over $40 million in suspicious transfers across at least 10 blockchains on 2 November, and the following day a Breachforums user advertised '700k Brazil crypto emails'; the two are unconnected on the evidence available and the advert may be recycled data.

    confidence 0.30Speculative
    Wrong if
    A sample shown to predate 2 November or to originate from a different Brazilian source.
    Watch
    A sample check and the exchange's updated statement.

Nothing to seeThe WAVLINK, Gitea, Apache Wicket and AI-agent-tooling CVE batches look like single-researcher audits and batch publication, not coordinated exploitation. The Telegram hacktivist traffic is mostly backfill with no event timing, and the ThreatFox volume and Mastodon chatter added no cross-source structure today.

Deep read: 672 items read

§4

Forecast

Dated, falsifiable calls. We mark our own homework in public.

Today’s calls

  1. DueWed 11 Nov7d

    CISA adds at least one further vulnerability to the KEV catalogue by 11 November 2026.

    Wrong ifNo new KEV entries appear in the CISA KEV feed between 4 and 11 November.

    f-2026-11-04-1

    Confidence0.85
  2. DueTue 10 Nov7d

    No single day between 4 and 10 November records more than 20 new thegentlemen victim claims across ransomware.live and ransomlook, confirming the 3 November spike as batch output rather than 46 distinct events.

    Wrong ifAny one day in that window shows more than 20 distinct new thegentlemen victims on the trackers.

    f-2026-11-04-2

    Confidence0.65
  3. DueSat 14 Nov10d

    Ubuntu moves the IPv4 TCP use-after-free kernel fixes out of its opt-in testing channel into regular updates for its three newest LTS releases by 14 November 2026.

    Wrong ifThe fixes remain testing-channel only, with no regular-update release, on 14 November.

    f-2026-11-04-3

    Confidence0.50

Scorecard

1 right · 2 wrong

  1. wrong

    CISA adds the FortiSIEM path traversal CVE-2026-106173 to the KEV catalog by 4 November, following the public proof-of-concept.

    Called Wed 21 Oct · due Wed 4 Nov

    Today is the due date; KEV on 2026-11-03 added only CVE-2026-119977 (WordPress), not CVE-2026-106173 (FortiSIEM path traversal). No FortiSIEM entry found.

  2. right

    NVD publication resumes in these feeds, with at least 20 new CVE entries ingested, by 6 November 2026.

    Called Tue 3 Nov · due Fri 6 Nov

    NVD items are present in today's feed starting 2026-11-01, well over 20 new CVE entries ingested before 2026-11-06 deadline.

  3. wrong

    The Gentlemen do not post another day of 20 or more leak-site items between now and 10 November 2026; the 2 November burst was a batch release, not a new rate.

    Called Tue 3 Nov · due Tue 10 Nov

    Tempo report shows The Gentlemen with 50 claims in 7 days (surging), and ransomlook shows a The Gentlemen leak posted on 2026-11-03, far exceeding 20 items in the window.

Glossary of terms you will be expected to nod at

Auto-update (n.)
A standing appointment at which you permit a stranger to write files to every one of your servers, agreed once, in a procurement meeting, by somebody who has since left.
§5

From our edge

Live attacks on our own servers in the last 24 hours. Yes, really ours.

Our own edge, last 24 hours: the Adversary DNA sensor profiled 64 hostile sources — 34 scanners, 27 bot crawlers, 2 brute forcers and one targeted operator — sending 9,687 requests between them, of which 23 went past reconnaissance into exploitation attempts, across 6,645 addresses from 99 countries, 2,127 of them seen for the first time today. Fourteen groups of addresses shared a behavioural fingerprint, including seven IPs on DNA dfef4501b216: 158.158.34.156 at hostility 6.7 sent 130 requests asking for /wp-content/plugins/hellopress/wp_filemanager.php and /this_is_a_new_hello_world.php, while 34.95.134.180 sent 277 requests at 207 a minute hunting /.git/config and /.env variants, and 207.175.85.46 managed 122 requests in about 1.0 seconds looking for /.aws/credentials.bak and /.docker/config.json. One address, 20.194.27.148, sent a single request shaped like CVE-2024-4577 exploitation at /index.php, which tells you precisely how long an exploit stays in the spray rotation regardless of what the patch notes say about the year. Raw logs never left the server; only these findings did.

64hostile sources profiled
9,687requests from them
23reached exploitation
14shared-DNA clusters
7.7/10top hostility
6,645addresses at our edge
2,127first seen today
99countries

The WTF list

  1. WTF #1

    The most determined visitor

    158.158.34.156, a scanner (hostility 6.7/10), sent 130 requests to one of our sites, probing for /wp-content/plugins/hellopress/wp_filemanager.php, /this_is_a_new_hello_world.php, /rapid.php, //44.php.

    Why it’s weirdHighest hostility on the board; its kill chain reached weaponization.

    attribution: unknownWed 4 Nov, 00:00 UTC

    Detail
    kill chain: reconnaissance → weaponization · DNA dfef4501b216
  2. WTF #2

    Same DNA, different addresses

    34.95.134.180, a scanner (hostility 6.3/10), sent 277 requests at 207 a minute to one of our sites, probing for /.git/config, /.env, /.env.local, /.env.production.

    Why it’s weird7 IPs share one behavioural fingerprint (DNA dfef4501b216): one tool or one operator rotating addresses: 158.158.34.156, 20.164.214.25, 3.142.247.125, 173.239.213.19, 34.95.134.180 and more.

    attribution: unknownWed 4 Nov, 00:00 UTC

    Detail
    kill chain: reconnaissance → weaponization · DNA dfef4501b216
  3. WTF #3

    Trying CVE-2024-4577 on us

    1 address sent 1 requests shaped like CVE-2024-4577 exploitation (/index.php), including 20.194.27.148.

    Why it’s weirdA known, named exploit still being sprayed at the internet; our edge saw it today, whatever the patch notes say about the year.

    attribution: unknown

    Detail
    probe family: webshell
  4. WTF #4

    The strangest shopping list

    20.194.45.183, a scanner (hostility 6.4/10), sent 46 requests to one of our sites, probing for /wk/index.php, /inputs.php, /ioxi-o.php, /function/function.php.

    Why it’s weirdAsked for files almost nobody else asks for: /wk/index.php, /inputs.php, /ioxi-o.php, /function/function.php, /rip.php.

    attribution: unknownWed 4 Nov, 00:00 UTC

    Detail
    kill chain: reconnaissance → weaponization · DNA 83562bd8e880
  5. WTF #5

    In a hurry

    207.175.85.46, a scanner (hostility 5.4/10), sent 122 requests to one of our sites, probing for /.aws/credentials.bak, /.aws/.env, /.docker/config.json, /.env-sample.

    Why it’s weirdFired 122 requests in about 1.0 seconds.

    attribution: unknownWed 4 Nov, 00:00 UTC

    Detail
    kill chain: reconnaissance → weaponization · DNA 18e39c4b41fb
§6

Tempo

Who is posting faster than their own normal. Computed from the Signal graph, not written by the editor.

231claims, 7 days
39of 271 groups active
thegentlemen50 this week · surging
qilin29 this week
N0n17 this week
  • Surgingthegentlemen — 50 claims in 7 days against 25.0 expected from its own prior rate (p < 0.001)
  • SurgingBooba Project — 11 claims in 7 days against 2.9 expected from its own prior rate (p < 0.001)
  • Surginggenesis — 9 claims in 7 days against 2.4 expected from its own prior rate (p < 0.001)
  • Surgingrhysida — 6 claims in 7 days against 1.5 expected from its own prior rate (p = 0.005)
  • SurgingSilentRansomGroup — 8 claims in 7 days against 2.8 expected from its own prior rate (p = 0.007)
  • Surgingsafepay — 10 claims in 7 days against 5.0 expected from its own prior rate (p = 0.031)

Gone quiet: gunra (61 days), lynx (67 days), AiLock (70 days), abyss (70 days).

Sector pressure, 30 days

  • UpTransportation — 2.1× its own prior share (70 claims, 34 expected)
  • UpEducation — 2.0× its own prior share (54 claims, 27 expected)
  • UpEnergy & Utilities — 1.4× its own prior share (36 claims, 27 expected)
  • DownAgriculture and Food Production — 0.7× its own prior share (22 claims, 34 expected)

25% of claims carry no sector label.

Brand churn

Of the 233 leak-site brands first seen since 2021-09-30, 76% were still claiming 30 days after their first post and 70% were still claiming 60 days after their first post and 68% were still claiming 90 days after their first post and 62% were still claiming 120 days after their first post (Kaplan-Meier; quiet = 60 days without a claim). 18% of the 214 old enough to judge posted on one day and never again; 39% made at least half their claims in their first week. New brands this month: 2 — newest: OBSOFT (2026-11-03, 6 claims), PixelLeaks (2026-11-03, 1 claim), everstake (2026-10-29, 7 claims).

Brands born before coverage began are excluded (their true age is unknown); a rebrand counts as one brand going quiet and another being born.

Targeting

Over the last 180 days, among brands with at least 20 labelled claims: safepay put 32% of its claims on Germany (6.2× every other brand's share); N0n put 33% of its claims on Transportation (7.4× every other brand's share); genesis put 95% of its claims on the United States (2.6× every other brand's share). Widest spread: nightspire across 10.6 effective sectors.

Lift = the brand's share of a sector or country over every other brand's share in the same window; exact binomial tails, Bonferroni-adjusted for the categories tested. Claims with no sector or an Unknown/Other label are excluded.

Posting clock

Over the last 180 days, among brands with at least 30 claims: clop released 94% of its claims on Wednesdays, half of them in 13:00–15:59 UTC; safepay released 63% of its claims on Mondays, half of them in 18:00–19:59 UTC; akira posted only 1 of 196 claims at weekends and half its claims in 12:00–13:59 UTC; half of play's claims appeared in 18:00–20:59 UTC. Leak sites overall are 32% quieter per weekend day than per weekday.

all brands by hour of day, 00–23 UTC; half of all claims land in 12:00–19:59 UTC.

Times are when trackers first saw each claim on the leak site, in UTC: not the attack time, not the operator's clock, and a site that is only reachable at certain hours shows the same pattern; batch releases show as spikes. Hour-of-day concentration: Rayleigh test; weekday shares: exact binomial against every other brand, Bonferroni-adjusted.

Leak-site claims as discovered by trackers, not confirmed breaches. Surprise compares the last 7 days with the group's own prior 173 days (Poisson). Victim claims: ransomware.live and leak-site trackers

Today’s numbers as JSON →

§7

Ride-alongs

Free, live tools. No sign-up, no sales call.

§8

Colophon

How this edition was made, plainly.

Edition 38 was made by machine and checked by code. Every item published by 15 public threat feeds in the 24 hours to 05:05 UK (672 items) was read in full: statistics computed in code, each slice read by Claude Sonnet, the structure synthesised by Claude Opus. Claude Opus then edited the paper. Every story cites the items it rests on; stories whose receipts did not resolve, or whose numbers were not in their sources, were dropped before publication (0 today). Forecasts are graded in public in later editions. 'From our edge' is live attack data from our own servers. Cost of this edition: $1.16. It is written by machine in the voice of ScottG, author of Everything Is Probably Fine, and nobody edited it by hand, which is either the point or the problem.

672items read
24hour window
$1.16cost of this edition
05:05 UTCgenerated

Sources this morning

  • ThreatFox160ok
  • Telegram124ok
  • NVD100ok
  • Signal Atlas95ok
  • Security news54ok
  • infosec.exchange38ok
  • ransomware.live30ok
  • RansomLook23ok
  • atlas22ok
  • RSS10ok
  • AlienVault OTX7ok
  • signal:ioc-family6ok
  • signal:kev1ok
  • CISA KEV1ok
  • Ransomware tempo · Ninja Signal graph1ok

ITEMS READ: 672SOURCES: 15STORIES DROPPED FOR LACK OF RECEIPTS: 0COST: $1.16STATUS: NOT ENTIRELY FINE

OWNER: the machineEDITOR: the rulesCALLS OPEN: 3NEXT REVIEW: tomorrow 06:00 UK