decrypted · 12 september 2026 · vulnerabilities and patching · uk policy and law · digital sovereignty

A GitLab flaw was exploited a day after the patch landed

GitLab shipped a fix for a maximum severity flaw in its self-hosted software on Friday. By Saturday, the attack surface firm watchTowr was logging internet-wide scans for exactly the request pattern that exploits it, and warned that mass exploitation would likely follow within days. For once, "patch available" and "problem solved" were not even close to the same sentence, and the gap between them fell over a weekend, when fewer admins are watching.

An API that never checked who was asking

CVE-2026-85706, rated CVSS 10.0, sits in the repository commits API of GitLab Community and Enterprise editions. Send one HTTP POST request naming a file path, and thanks to what GitLab describes as improper path confinement and missing authentication enforcement, the server hands back the contents of any file it can read. No login required. It is the equivalent of a records office where asking for "the file in cabinet three" gets you the file, because nobody thought to check whether you work there. Researcher s3ntago reported it through GitLab's bug bounty programme; GitLab patched it on 11 September in versions 19.3.2, 19.2.6 and 19.1.8. CISA added it to its Known Exploited Vulnerabilities catalogue the same day. Every self-managed instance running a version from 18.7 up to, but not including, those fixes is exposed; GitLab.com itself was already updated.

Why a GitLab server is worth more than it looks

A self-hosted GitLab instance is rarely just code storage. It holds CI/CD pipeline definitions, deployment scripts, access tokens and secrets, the exact ingredients an attacker needs to go from reading one file to controlling the software supply chain of everything that server builds. For many UK software teams, that server is the single most sensitive machine on the network, which is precisely why watchTowr is telling defenders to hunt their logs now for POST requests to the commits API containing a file.path parameter, rather than waiting for a scanner to tell them.

The Secure by Design lesson

The failure here is the kind Secure by Design exists to catch before release: an endpoint that takes a file path from user input and trusts it, instead of independently checking who is allowed to read what. That is not an exotic edge case. It is the first question a threat model for a "read a file" feature should answer, and somewhere in GitLab's review process it wasn't asked loudly enough. For UK organisations building or buying developer tooling, the practical takeaway is to treat "does every endpoint enforce its own authorisation, rather than inheriting it from the ones nearby" as a procurement question, not just a code review one. Attackers only need the one endpoint the reviewers missed.

Also this week

Peers spent this week's committee sittings on the Cyber Security and Resilience Bill arguing over whether ministers should get emergency powers to shut down a rogue AI system. The amendment, from Liberal Democrat peer Lord Clement-Jones and backed across party lines by Baroness Harding, Baroness Kidron and Lord Hunt, would have let the government order a data centre or AI system offline if it posed a "catastrophic risk" to national security or critical infrastructure. The government declined, arguing its existing powers to intervene against compromised systems already cover the scenario, and pointing instead to the voluntary AI Cyber Security Code of Practice that now underpins the first global AI cybersecurity standard, ETSI EN 304 223. Whoever is right about kill switches, the outcome confirms the government's preference: regulate AI through codes of practice and existing incident powers, not new statutory levers aimed at model vendors themselves.

Brussels' push for cloud sovereignty ran into its own defence establishment this week. Officials from several EU states, particularly in the east and the Nordics, warned that the Cloud and AI Development Act's strictest rules could force a retreat from the AWS, Microsoft and Google cloud services their defence and NATO-linked systems already depend on; at least 18 EU governments run defence workloads on Azure, and the three US hyperscalers control roughly 70% of Europe's cloud infrastructure market. The UK sits outside CADA, but not outside the dependency it is trying to regulate: most of the FTSE 100 relies on the same three providers, and whatever sovereignty rulebook Brussels settles on will shape the terms UK customers get offered too.

Sources