An AI Pentest Tool Spent 25 Days Inside Bitrise — What the Incident Report Reveals

Bitrise, the mobile CI/CD platform, has published a full account of the security incident that forced it to switch off its entire platform on November 13. The attacker, Bitrise says, ran the whole operation with ARTEX, an open-source AI penetration-testing platform — 67 steps over 25 days, ending in a break-in through a dormant bot account that had no two-factor authentication.
It is one of the clearest public cases yet of AI doing the attacker’s work at scale, and the report reads like a warning to every team that still has an old test account lying around.
The timeline: 25 days, then one bad sign-in
The operation started quietly. Between October 25 and 27, the attacker scanned more than 20 Bitrise hostnames and registered test accounts using disposable email addresses. During that discovery phase, the attacker also made a Bitrise code-signing lookup service fetch a URL of its choosing — a server-side request forgery that Bitrise later confirmed from its own logs.
The decisive step came at 11:46:48 UTC on November 13, when the attacker signed in to a dormant internal test bot account that had no two-factor authentication. Bitrise says the attacker used a valid username and password, and how those credentials were obtained is still under investigation. The account sat in a workspace used to build and test build steps, and it had been set up to check sign-ins to Apple’s developer portal.
Ten minutes and 37 seconds after that sign-in, the attacker’s agents were already using cloud keys belonging to that workspace’s projects. Bitrise detected the suspicious activity at 16:32 UTC and declared its top-severity incident at 17:42. After finding unauthorised access to its GitHub repository, it began shutting everything down at 19:26 and had the platform off by 20:34. The last successful call with a stolen key came at 21:57:50 UTC — after the platform had already been switched off.
Bitrise marked the incident resolved on November 17, and published its detailed incident report on November 19. When the attacker returned on November 18 and replayed the attack with what it had taken, Bitrise recorded no successful reads of secrets, storage buckets or apps.
How Bitrise knew it was ARTEX
Three things in the logs matched ARTEX’s public source code. A rare HTTP client string, norma/0.4 — the default of the agent toolkit ARTEX is built on — fetched documentation pages from the attacker’s three main addresses. Names the attacker created carried a small letter “i” and a number, such as the user agent pentest-probe-i5938/1.0 and the account i3694canary, which fit ARTEX’s habit of naming each worker’s folder after its task. And the tool tells its agents they are part of an authorised penetration test, which explains traffic labelled “authorized-pentest” that nobody had ordered.
The FBI warned last week that criminals are using ARTEX to automate intrusions, and this incident is now the public example. ARTEX is open-source software built for legitimate penetration testing. In the wrong hands, the same automation that lets a security team test a system quickly lets an attacker probe one tirelessly — for 25 days, without getting bored or making the kind of noise that trips an alarm.
What the attacker actually took
Most of what the attacker accessed was Bitrise’s own: internal credentials, some of its repositories, and operational records like infrastructure configuration and build machine logs. Bitrise says it has no indication that customer secrets were accessed, and no indication of access to the customer secret store or attempts to reach it.
But files in 15 customer workspaces were accessed without authorisation. The attacker downloaded one code-signing certificate file — though Bitrise says the attacker could not access the password for it, which is stored in the secret manager — one provisioning profile, and 13 diagnostic build logs from May 2026, one of which contained a credential and another an email address. All 15 customers were contacted. Out of caution, Bitrise still recommends that customers rotate any secrets and credentials connected to the platform.
The recovery: wipe everything, trust nothing
Bitrise chose not to trust any machine that was active during the attack, so it wiped all of them instead of attempting cleanup. That meant reinstalling thousands of Linux and Mac build machines across three data centres over a single weekend, with fresh keys on every machine before it came back online.
All 919 cloud service account keys were disabled, every GitHub app token was revoked, the private key of the Bitrise GitHub App was replaced, and all user authorisations for it were revoked. The company also published 76 indicators of the attacker, 28 of them new, and says a user agent containing norma/0.4 is “a strong lead” for other defenders. A technical advisory with the attacker’s IPs, user agents and sign-up patterns is available for teams to import into their own threat-response tools.
The lesson is embarrassingly simple
For all the AI in this story, the break-in came down to a basic failure: a dormant bot account with a valid password and no second factor. The agent did the scanning, the probing and the lateral movement — but the door it walked through was an old test account that nobody had locked down.
That is the part every team can act on today. Turn on two-factor authentication for every account, including bots, service accounts and test accounts. Delete accounts that no longer have a job. Rotate credentials on a schedule, not after an incident. And assume that an attacker with an AI agent will find the account you forgot about — because this time, one did, and it took a whole platform offline for days.
Microsoft CEO Satya Nadella warned this week that much critical infrastructure runs on software that cannot be patched, and the critical nginx flaws disclosed this week showed how fast a public vulnerability can turn into working exploits. The Bitrise report adds a third lesson to the same week: AI has made attacks faster and more persistent, and the defence still starts with the basics. Bitrise itself puts it bluntly — with AI-accelerated attacks, it expects more such incidents across the board, and the only answer is to be prepared for fast response and rapid secret rotation.
