The machine that caught this week's first GitLab attack was never a real one

The alarm came from a machine nobody uses

GitLab, the code-hosting platform thousands of companies run themselves, shipped an emergency fix on Monday 17 August [1][2]. The flaw is rated 9.4 out of 10 [1]. It let a stranger with no account and no password change or delete public projects and user data [1][2].

The security firm WatchTowr says its honeypot network caught the first real-world attempts to use it [1]. A honeypot is a machine put on the internet as bait. It runs no real work and serves no real users. Nobody has any legitimate reason to touch it, so anything that does is an intruder by definition.

That is the whole trick. There is no judgement call, no “was this normal” - the alarm is the answer.

The gap was about two days from public disclosure to attack [1]. WatchTowr says it reproduced the flaw “within minutes” of disclosure, armed only with the advisory and the patch [1]. No public exploit code had been released [1]. Its principal researcher Jake Knott calls that narrow window the new reality [1].

What worries researchers most is not deletion. The flaw can forge merge records - the approval stamp that says a human reviewed a code change [1]. Mondoo co-founder Patrick Munch notes that in every supply-chain worm tracked this year, attackers had to work around code review [1]. This does it for free.

Anyone running their own GitLab should update to 19.2.4, 19.1.6, 19.0.8 or 18.11.11 [1]. WatchTowr suggests searching web logs for requests containing @gl_introduced [1].

Four more doors, and how long each stayed open

Zimbra, an email suite used by hundreds of millions of people, is under active attack [3]. Poland’s national cyber-response team says it has seen attacks this week on a flaw patched back on 20 July [3][4]. Shadowserver, an internet-scanning watchdog, counts more than 12,100 Zimbra servers still reachable online [3]. It cannot say how many are patched - or how many are decoys [3].

Citrix patched a 9.3-rated flaw in NetScaler on Wednesday [5][6]. NetScaler sits at the public edge of corporate networks and handles remote access, so it is a standing target. Rapid7 expects attackers to move on it shortly for exactly that reason [5].

CISA, the US cyber-defence agency, added two flaws in TrueConf Server, a video-conferencing product, to its must-patch list on Thursday, citing evidence of active use [7].

And researchers at WatchTowr say attackers are using a flaw in MLflow, an open-source tool for running AI models [10]. It lets them reach the internal service that hands a cloud server its own keys, and take them [10].

Sometimes there is nothing left to patch

CISA, the FBI and the US health department refreshed their advisory on the Medusa ransomware gang on Tuesday [8][9]. The original went out in March 2025; this is a year of further investigation folded in [8][9]. Confirmed victims rose from more than 300 to more than 500 as of April 2026 [8].

One line is worth sitting with. The agencies say Medusa exploits “newly announced exploits within 24 hours” - and has been seen using them up to a week before the flaw was publicly disclosed [8].

If that is right, patching cannot be the whole answer, because for part of that week no patch existed. Something has to notice the intruder instead.

The intruder built to look like your own software

Which is exactly what one new toolkit is designed to defeat. Researchers at Ontinue disclosed TWINLOOT, a hidden program that runs its entire remote-control channel through Microsoft’s own services [11].

Instructions are left as files in SharePoint Online. Live operator access is relayed through Microsoft Teams servers [11]. The traffic is driven by a headless copy of the victim’s own Edge browser, which Ontinue says makes it “virtually indistinguishable from legitimate network activity” [11]. It also shows fake lock screens to harvest Windows passwords [11].

Google reported the same instinct in three suspected Russian spying groups, tracked as UNC6293, UNC7005 and UNC5976 [12][13]. They have moved to abusing genuine sign-in flows - the real Google screens that ask you to let an app in [12][13]. Targets are academics, aerospace, defence, government and think tanks in Europe and the US [12][13]. Each campaign hit fewer than 100 people, with under 10 actually compromised [12].

Google’s point is blunt: because the sign-in page is real, victims may not recognise it as phishing at all [12].

Still, mostly, a password

The security firm Huntress recorded a 155-fold rise in password-spraying attacks in the first half of 2026 [14]. Password spraying tries one common password against thousands of accounts, slowly, to stay under lockout limits. In mid-June alone Huntress logged over 81 million login attempts from one campaign against a Microsoft admin tool [14]. It ended in 78 account takeovers in a fortnight [14]. The figures are Huntress’s own, published on BleepingComputer [14].

The human version is smaller and sharper. A Huntress researcher was approached on X after Black Hat and Def Con by someone posing as a senior CoinDesk marketing executive, asking for help with a made-up conference [15]. The researcher did not fall for it, so the same person came back the next day with a second lure, then a third pretext [15]. If you attend a big industry event, expect the follow-up messages to include a few that are not real [15].

Poisoned before the code ever ran

Someone took over the account behind arrayref, a widely used building block for Rust programs, and pushed a booby-trapped release [16][17]. Two sibling packages, internment and append-only-vec, went the same way inside a 23-minute window [16]. The Rust project pulled all three, 86 to 107 minutes after publication [17].

The nasty part is when it ran. The malicious code sat in a build script, so simply compiling a project that used it was enough [16][17]. Nothing had to call the library at all [17].

Download figures are counted two ways here. BleepingComputer puts arrayref alone at more than 53 million downloads over 90 days [16]. The Hacker News totals 245 million across all three packages [17]. There is no fix and no CVE number - the industry’s standard label for a known flaw - so developers are told to pin arrayref at 0.3.9 or earlier [17].

The same week, 16 packages were planted in RubyGems, the package library Ruby programmers pull from, under near-miss names like brumdler and activesupmport [18]. They steal browser logins, crypto wallets and seed phrases [18]. The trick is nothing more than a typo: names one slip of a finger away from real ones. All 16 have been pulled [18].

A quieter week for scheduled patches it was not. Atlassian and Splunk together shipped fixes for over 250 flaws [19]. Cisco patched 15, four of them critical in its Crosswork software. Three of those four carry the maximum rating, 10 out of 10 [20].

14,000 cameras, and the folder the attacker left open

The under-covered story is a mirror of the lead. Researchers at Hunt.io say one operator compromised more than 14,530 Dahua internet cameras across Ukraine and Russia between 17 June and 22 July [21]. Most fell to guessed passwords [21].

On 1,923 of them the attacker planted a second account, username and password p2pwn and p2password [21]. It is stored apart from the admin login. It survives a password change, and on most firmware it survives a factory reset [21].

Hunt.io found all this because the operator left their working folder open on a web server - 407 MB of logs, credentials and captured images, 2,616 files in all [21]. The defenders’ tripwire caught GitLab. Here the attacker built their own.

Elsewhere: Bitdefender detailed SilkParasite, a China-linked spying operation that ran for nearly a year against government bodies across Central Asia, using seven malware families [22]. And Australian aparthotel chain Quest told guests their names and contact details leaked through a third-party database provider, with dates of birth in a small number of records [23].