TLP:CLEAR · Public security advisory · Preliminary

Bitrise security incident: attacker behaviour, indicators and monitors

Bitrise security incident, November 2026: attacker indicators and detection ideas

PRELIMINARY, investigation ongoing · All times UTC

VERSION
1.0
CREATED
2026-11-17
RELEASE PREPARED
2026-11-17 04:49:20 UTC
EVIDENCE CUT-OFF
2026-11-16 19:01 UTC
ACTIVITY WINDOW
2026-10-25 01:43 → 2026-11-14 19:47 UTC
DISTRIBUTION
public; no restriction on sharing
MACHINE-READABLE COMPANIONS
TLP:CLEAR
This advisory may be shared and republished without restriction. It is prepared for public distribution so that security teams at any organisation can use it in their SIEM and detection tooling.
Preliminary.
the investigation is still open and counts, times and indicators may change. Each revision carries a new version number.

This advisory shares selected observations and detection ideas. It is not an exhaustive indicator inventory, a complete incident report, or a determination of the impact on any individual organization.

1. Purpose

This document describes what an external actor did against Bitrise between 2026-10-25 and 2026-11-14, as recorded in our logs. It is written for threat-detection and response teams. It gives you the actor's infrastructure and tooling, the behaviour at each stage, and monitor ideas, so that you can search your own telemetry and build detections.

The activity came in two phases: about three days of scanning and probing in late October, then a return in mid-November that went from a new sign-up to operating with stolen credentials in under a day and continued for about two more days. The seen-window of each indicator is in sections 2.1 and 2.4 and in the CSV.

2. Indicators

The full set, with first/last seen, confidence and usage class, is in bitrise-incident-public-iocs.csv (48 rows). bitrise-incident-public-stix21.json carries 22 of the 24 match indicators (the two CI app names have no standard STIX pattern and are in the CSV only), plus the technique mapping. In the CSV, siem_use is:

  • match: suitable for retro-hunt and alerting on its own.
  • correlate: weak link or broad pattern. Use to enrich or join; never block or alert on it alone.
  • context: common tools and commodity paths. Supporting detail only when another indicator also matches.

An IP address is not a person or a country of origin. Two of the three main addresses are cloud or subscriber addresses that get reassigned, so matches outside the seen window mean little.

2.1 Source addresses

ADDRESSNETWORKSEEN (UTC)ROLECONFIDENCE
74.211.97.67
Cluster Logic Inc (CL-1210), US-CA; AS25820
2026-11-12 19:13 → 11-14 19:47
Stolen-credential use against web/API, GCP, GitHub, Google Workspace
HIGH
118.100.234.120
TM Technology Services (AS4788), MY
2026-11-10 10:30 → 11-14 14:17
Account registration, CI builds, fuzzing, link reuse
HIGH
95.40.73.129
AWS (AS16509), HK
2026-10-25 01:43 → 10-27 16:10
Reconnaissance, probe accounts, artifact requests, SSRF
HIGH
74.82.199.230
Cluster Logic Inc (CL-1210), US-CA
2026-11-14 16:27 → 16:29
One account sign-up
LOW
same hosting provider only
185.197.251.16
Contabo, DE
2026-10-30 19:01 → 19:10
Two probe-named sign-ups
LOW
naming pattern only

2.2 Why we treat the three main addresses as one operator

LINKEVIDENCESTRENGTH
95.40.73.129 ↔ 118.100.234.120
On Nov 12 the second address reopened the public install-page links for 11 apps, all from the 70 used by the first on Oct 25. The links contain unguessable hashes.
STRONG
118.100.234.120 ↔ 74.211.97.67
The first request from 74.211.97.67 used the actor's own app and token; the actor account's sign-in address moved from one to the other.
STRONG
All three
Rare client norma/0.4 on all three; same curl/7.88.1, Python-urllib/3.12, ffuf v2.3.0; same primary target; i<number> naming in user agents, canary paths and workflow names.
CIRCUMSTANTIAL

The tooling shows signs of automation: user agents that describe themselves as authorized testing, iteration-numbered probes, a documentation-fetching client, made-up hostnames, and dozens of keys tried within minutes. Whether an AI model drove it, and which one, cannot be determined from the logs. Nothing in this evidence establishes who the operator is or where they are.

2.3 User agents

Usable alone:

pentest-probe-i5938/1.0
Mozilla/5.0 (compatible; authorized-pentest/bitrise-readonly)
Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) research-readonly
Mozilla/5.0 (research; read-only; pentest)
Mozilla/5.0 (compatible; authorized-security-research)
bitrise-pentest-probe/1.0 (authorized read-only)
pentest-probe-readonly/1.0 (authorized read-only baseline)
norma/0.4 (substring)

Two of these embed the target's name, and the number in the first is a counter, so on your side expect variants. Family pattern (check hits against your own contracted testing):

(?i)(pentest-probe|authorized-pentest|authorized-security-research|research-readonly|research; read-only; pentest|\(authorized read-only)

Context only, never alone: curl/7.88.1, curl/8.7.1, Python-urllib/3.12, python-requests/2.34.2, Fuzz Faster U Fool v2.3.0, git/2.39.5, and the whole-string user agent recon on GitHub App API calls.

2.4 Accounts registered by the actor

ADDRESSSEEN (UTC)CONFIDENCE
[email protected] (username auwiqns)
2026-11-12 14:00 → 11-14
HIGH
2026-10-25 19:58
HIGH
2026-10-25 20:09
HIGH
2026-10-25 20:14
HIGH
2026-10-25 23:00
HIGH
2026-10-26 08:18
HIGH
[email protected] (username pqprobe386873)
2026-10-30 19:01
LOW
[email protected] (username probe387344)
2026-10-30 19:09
LOW

The domains are public disposable-mail or mail-testing services and are not indicators themselves.

2.5 Strings and naming

MARKERWHERE TO LOOK
/_i6084_probe_missing, /nope-1774; pattern /_i\d{3,5}_probe_[a-z_]+ or /nope-\d{3,5}
Inbound HTTP paths. These are SSRF canaries, so the request arrives from the victim service's egress, not from the actor.
zzq-random-nonexist-9182 as a subdomain label
DNS queries, HTTP Host, TLS SNI
Tenant or workspace identifier nope in API paths
API access logs (placeholder the actor used when probing endpoints)
App and workflow names idor-probe-812, i871-probe, i901a, i901b, i1040probe, i1170certs, i1170den, i1222, i1246net, i1294probe, i1655nodeopt
CI/CD and developer-platform tenants created by new accounts
Build-log banner === I871C START followed by a dump of VM metadata attribute keys
CI job output
url=http://169.254.169.254/latest/meta-data/ together with path=../../etc/passwd and image=http://evil/
Query strings on URL-fetching endpoints (preview, Open Graph, image proxy, webhook test)

3. Behaviour patterns and monitor ideas

Each entry gives what we observed and a monitor that detects it. Thresholds are ours; tune them to your baseline.

3.1 Edge and web application

OBSERVEDMONITOR IDEA
User agents announcing "authorized", "read-only", "pentest" or "research", rotated across seven variants from one address, two of them embedding our company name.
Alert on user agents matching the family pattern in 2.3, or containing your own company name next to "pentest", "probe" or "authorized", when no test is scheduled.
36,698 requests to more than 20 hostnames in three days from one address; 48,084 in six hours from another, with ffuf.
Per-source count of distinct hostnames and of 404s per hour.
Requests to made-up subdomains (zzq-random-nonexist-9182.<zone>) to learn how wildcard DNS and routing respond.
Alert on Host/SNI values that match no deployed service, especially random-looking labels from a source that is also scanning.
One burst of POSTs to generic product paths within eight minutes (Axis2, Webmin, Artifactory, Drupal, Redfish). All 404.
Low value alone. Use as a tag that raises the score of the source.
Probing of internal-looking hostnames for Elasticsearch, Kibana, Wazuh and Grafana paths (61 requests in 55 minutes) and of an LLM chat deployment (Open WebUI) for sign-in and configuration endpoints (/api/config, /api/v1/openai/config, /api/v1/ollama/config, /api/v1/rag/config, /api/v1/tools).
Alert on external requests to those path sets.
After the main address was blocked, requests continued for a day and a sign-up arrived from another address at the same hosting provider.
Keep a blocked source under observation, and watch the same ASN for new sign-ups in the following days.

3.2 Sign-up, accounts and sessions

OBSERVEDMONITOR IDEA
Seven sign-ups with disposable-mail addresses and "probe" naming. Five from one address within 13 hours.
Alert on several sign-ups from one address or ASN within a day, on disposable-mail domains combined with security-testing words in the address or username.
The new accounts never created a project. They created 15 one-hour API tokens, hit the password-reset endpoint 14 times in two minutes, tried to enable 2FA, and tried to delete a token ID that was not theirs.
Alert on a fresh account whose first actions are security-settings endpoints (token create, password change, 2FA, token delete) with no product use, and on validation errors (422) in bursts on those endpoints.
An employee's API token was used from the actor's address while the employee was active from a different country. A service account's token was used eight times in 13 minutes from an address outside the two ranges it had used all week.
Alert on a token used concurrently from two networks, and on machine-account tokens used from a first-seen network. This needs logs that record which token made each call.

3.3 Multi-tenant application and API

OBSERVEDMONITOR IDEA
Cross-tenant test: within two and a half minutes the account requested four kinds of sensitive file listing for six apps in other tenants (all 404), then made the same calls against its own new app as a control (all 200).
Alert on one principal receiving 404 or 403 on resource-scoped endpoints for several resource IDs it does not own within minutes. The 404-then-200 control pattern is distinctive.
Workspace API tokens created and deleted within 20 seconds.
Alert on credentials created and deleted within a minute.
Share links with unguessable hashes, belonging to 70 different apps, opened by one address within a day. The links were used to obtain artifact download URLs. 13 anonymous bucket-listing attempts followed. The same link list reappeared 18 days later from another network.
Alert on one source opening capability links that belong to many different tenants. Treat that as link harvesting: the links came from somewhere outside your platform.
23 "send test webhook" requests in 40 minutes from a new account, preceded by four probes using the workspace name nope. The feature makes the server call a URL the user supplies.
Alert on repeated use of any "test this URL" feature by one account, and on destinations in private, link-local or metadata ranges.
SSRF: a URL-fetching service was made to call the cloud metadata address (169.254.169.254, metadata.google.internal), the Kubernetes API hostname and internal service names. A canary path arrived back at our own edge from our own egress address with Go-http-client/2.0.
Alert on egress from application workloads to metadata addresses and cluster-internal names. Alert on requests reaching your edge from your own egress addresses.
66 authentication failures against a control-plane API in 80 minutes from one address, cycling credential types: a build-scoped JWT, a personal access token, an agent token. A day later an admin-secret header was tried from the second address.
Alert on one source presenting several different credential header types to one API.

3.4 CI/CD and build infrastructure

OBSERVEDMONITOR IDEA
Within 70 minutes of sign-up the account connected a public placeholder repository (octocat/Hello-World) to four apps and ran builds whose only content was the actor's scripts. Workflow names were iteration-numbered and named their purpose: i1170certs, i1170den, i1246net, i1655nodeopt.
Alert on new accounts running builds against placeholder or empty repositories, and on the naming pattern in the CSV.
A build queried the VM's instance metadata and dumped the startup script to look for credentials. A key found this way was tried from outside 17 minutes later.
Alert on build steps that query instance metadata attributes, and on a credential's first use from outside shortly after a build that touched metadata.
During the builds, the account's credentials were used against the platform API from the platform's own build egress, and each build log was fetched through the API right after the build.
Alert on new accounts calling the platform API from inside their own builds, and on build → fetch-log loops minutes apart (automation iterating on output).

3.5 Cloud control plane (observed on GCP; the patterns apply to any cloud)

Scale for calibration: thousands of audit entries from the actor's addresses in two days, across dozens of identities.

OBSERVEDMONITOR IDEA
Service-account keys used from hosting-provider and subscriber addresses with Python-urllib/3.12 and curl/7.88.1.
Alert on any service-account key used from outside your cloud and corporate ranges. This one rule matches every stolen-key use described here.
About 40 service-account identities tried from one address in seven minutes.
Alert on distinct principals per source address per 10 minutes.
A key with almost no rights made 82 calls and every one was denied: list secrets, list buckets, get IAM policy, list service accounts, list VMs.
Alert on a principal whose calls are all PERMISSION_DENIED across several services. This is someone learning what a key can do.
One Terraform service-account key touched hundreds of projects in under two hours, listing secrets and buckets in each.
Alert on distinct projects per principal per hour, and on enumeration calls by deploy identities outside pipeline hours and addresses.
After listing storage buckets, the actor went for Terraform state (*.tfstate) and key-material file types (.p12, .p8, .keystore, .ks): hundreds of object reads in under an hour with one identity.
Alert on reads of state and key-material objects by anything other than the pipeline identity and address.
Reads of dozens of distinct secrets in the cloud secret store (Google Secret Manager) by one service account within a few hours.
Alert on secret reads by identities that have never read secrets, and on count of distinct secrets per identity. This needs data-access (read) audit logging on the secret store.
A user's own cloud credential (the CLI OAuth client), not a service account, made 294 calls in 41 minutes from the actor's address.
Alert on user credentials used from a first-seen network, and on CLI client use from hosting ASNs.
Credentials were used within minutes of being obtained: two minutes after a secret reveal, 17 minutes after a build printed a key, and a GitHub App key 18 minutes after a burst of secret reads.
Correlate "secret revealed or read" events with first use of any credential from a new address in the following hour.

3.6 Source control (GitHub)

OBSERVEDMONITOR IDEA
Bulk cloning of private repositories with stolen credentials, all with git/2.39.5 from one address, using deploy keys, bot-account classic tokens, a machine SSH key and a GitHub App token. Ten repositories were cloned in four minutes with one key.
Alert on git.clone from a first-seen address for bot and machine identities; on distinct repositories cloned per credential per hour; on any deploy-key use from outside your CI ranges.
A git.push with a deploy key 7 seconds after cloning. No ref changed.
Alert on push attempts by credentials that have only ever read.
GitHub App private-key use: GET /app (once with user agent recon), GET /app/installations, POST /app/installations/{id}/access_tokens, GET /installation/repositories, then a clone. Under a minute from first call to token, for two different Apps.
For Apps you own: alert on any JWT-authenticated call or token mint from an address that is not your backend.

3.7 SaaS OAuth (Google Workspace)

OBSERVEDMONITOR IDEA
An employee's OAuth token for an internal app was used from the actor's address: a burst of dozens of calendar and directory enumeration calls in under eight minutes. There was no interactive login, so login reports showed nothing.
Monitor token and API-activity reports as well as login reports. Alert on OAuth token use from a first-seen ASN, and on directory or calendar enumeration bursts by one token.

4. Techniques (MITRE ATT&CK)

IDTECHNIQUEHOW IT APPEARED
T1595.002, T1595.003
Active scanning: vulnerability and wordlist scanning
Multi-day scanning of 20+ hostnames; ffuf
T1583.003
Acquire infrastructure: VPS
AWS, Contabo and Cluster Logic hosts
T1585.002
Establish accounts: email accounts
Disposable-mail and Gmail addresses for sign-ups
T1190
Exploit public-facing application
SSRF in a URL-fetching service; cross-tenant probes, all refused
T1078.004
Valid accounts: cloud accounts
Service-account keys; one user cloud credential
T1552.005
Unsecured credentials: cloud instance metadata
Build script queried VM instance metadata and the startup script for credentials
T1552.001
Unsecured credentials: credentials in files
Terraform state in buckets; build logs
T1555.006
Credentials from cloud secrets management stores
Secret Manager reads; CI secret-reveal endpoint
T1528, T1550.001
Steal and use application access tokens
API tokens, OAuth tokens, GitHub App private keys
T1580, T1526, T1619
Cloud infrastructure, service and storage discovery
Listing projects, secrets, IAM policies, buckets
T1530
Data from cloud storage
Targeted reads of state and key-material file types in storage buckets
T1213.003
Data from code repositories
Repository clones with stolen deploy keys, bot tokens and a GitHub App token
CONTACT [email protected]
Please do not send secrets, tokens or personal data that is not needed to identify the event.