Перейти к содержанию
Magpie holding a silver key on weathered copper pipes at misty dawn
Новости

Weekly briefing, 17 November 2026: CI secret stealer hits 1,806 repos, 1,000-flaw Patch Tuesday, AWS outage

Коротко

The wfx-loot campaign used 4,837 stolen GitHub accounts to steal CI secrets and cloud roles from at least 1,806 repositories. Also: a 1,000+ CVE Patch Tuesday, AWS us-east-1 sign-in outage, Expedia breach and Anthropic's IPO.

16 мин чтения

The week in brief

This was a week for the people who run build pipelines and cloud accounts. On 10 November, an attacker began using thousands of stolen GitHub accounts to push a credential stealer into repositories those accounts could write to. By 12 November, security researchers had confirmed the stealer had run in at least 1,806 of them. It is the main story this week and gets the full analysis below, because it goes after the part of a software company that usually gets the least review: CI workflows and the cloud roles they can assume.

On the same day, Microsoft shipped one of the largest Patch Tuesdays ever recorded. It contained more than 1,000 CVEs and one exploited Outlook zero-day, which a researcher found using Claude. AWS's oldest region lost IAM Identity Center sign-ins for part of Sunday morning. Expedia confirmed a breach through a third-party data platform. Anthropic priced its IPO for this week, and Stripe closed its purchase of OpenRouter. In the EU, the GDPR side of the Digital Omnibus moved forward in Parliament committee. In the UK, Ofcom showed that giving a regulator inaccurate answers carries its own fine.

Last week's briefing on AI-built Linux exploits and the Conductor deadline is the background to several of these stories. AI-assisted vulnerability discovery is now affecting patch volumes in production.

Dates to put in the diary

DateWhat happensWho should care
18–19 Nov 2026Anthropic prices its IPO (18th) and starts trading on Nasdaq as ANT (19th) (Yahoo Finance)Teams with Claude in their product or toolchain
24 Nov 2026CISA deadline for US federal agencies to fix the Outlook zero-day CVE-2026-124194 (SOC Prime)Anyone running classic Outlook on Windows
2 Dec 2026GitHub blocks `pull_request_target` by default in public repositories (summary of GitHub's changelog)Open-source maintainers and any public repos with Actions
2 Dec 2026EU AI Act transparency deadline (our guide)SaaS products with chatbots or generated content
14 Dec 2026GitHub retires six Copilot models (Make Tech Easier)Teams that pin Copilot models
Upcoming dates to put in the diary

The roundup

1. Microsoft's November Patch Tuesday: more than 1,000 CVEs and an exploited Outlook zero-day found with Claude

Microsoft patched 1,093 CVEs on 10 November: 63 critical, 1,029 important and one moderate, across 287 products and components. The only one exploited in the wild is CVE-2026-124194 in classic Outlook. Opening a crafted file makes Outlook leak the user's NTLMv2 hash, which can then be relayed or cracked. Tenable notes that Microsoft's advisory credits a researcher who used Claude to find it. CISA added it to its Known Exploited Vulnerabilities catalogue the same day.

Aerial view of endless rows of apple crates in an autumn orchard
More than 1,000 fixes in one release: triage only works with inventory and automation.

Rapid7 observed that "until a few months ago, these volumes would have been unimaginable". It also noted that the number of vulnerabilities known to be exploited has not risen alongside them.

For software teams, the developer tooling matters most. The patched products include Visual Studio Code (an 8.6 RCE and an 8.8 security feature bypass), GitHub Copilot, Azure Pipelines, the MSAL libraries for Node.js and JavaScript, and Microsoft's MCP Gateway. Rapid7 also flags that .NET 8 LTS and PowerShell 7.4 LTS reached end of support this week. If either is still in production, moving off it is now security work, not tidying up.

Sources counted this release differently:

SourceCount reportedWhat it covers
Tenable1,09363 critical, 1,029 important, 1 moderate
Rapid71,092Microsoft's own products only, 745 of them in Windows
SOCRadar1,094Calls it the second-largest monthly release on record
BleepingComputer1,235Calls it a record; its list includes Chromium-based Edge fixes
CrowdStrike1,252Counts 68 critical
How different sources counted November's Patch Tuesday

By type, Cyber Security News counts 535 elevation-of-privilege flaws, 296 remote code execution, 156 information disclosure, 46 denial of service, 27 security feature bypass, 22 spoofing and 11 tampering. The point for buyers is that nobody triages a thousand CVEs by hand. You need an asset inventory and automatic patch rings, so that you only have to decide about the exceptions.

2. AWS US-EAST-1 loses IAM Identity Center sign-ins for part of Sunday

On 15 November, problems started at 3:05 AM PST in AWS's Northern Virginia region. Customers using IAM Identity Center could not sign in to the console or to dependent services, and AWS Sign Up failed, according to AWS's incident updates as reported by GBHackers. For a while, affected customers had no workaround.

Engineers mitigated the authentication problem at 5:20 AM PST, and the final all-clear came at 6:53 AM PST. Eleven services were affected: IAM Identity Center, IoT Device Management, NAT Gateway, Sign Up, Step Functions, CloudWatch, Connect, Elastic Load Balancing, OpenSearch Service, VPC Lattice and Kiro (Cyberpress). AWS did not attribute the outage to malicious activity.

The lesson is the same as in earlier us-east-1 incidents. Your workloads may run elsewhere, but your way into the cloud often depends on one region. Check whether you have a tested break-glass path into production that does not go through Identity Center.

3. Expedia confirms a breach through a third-party data platform

On 12 November, Expedia Group confirmed that attackers reached its systems through a compromised third-party data platform provider, Skift reported. Expedia did not name the vendor or say what data was exposed.

Separately, Wiz and Sentra documented an attack between 14 and 26 October on the cloud data platform Qubole, a former Expedia vendor. Sentra found that former customers had never revoked Qubole's access permissions. Expedia has not confirmed any link to Qubole. The extortion group n0n claims more than 830 million loyalty records, 58 million traveller profiles and 333 million consumer account records, per SC World. These are attacker claims and have not been verified.

The practical lesson is to review vendor access at the end of a contract, not just at the start. Remove IAM roles, service accounts and data-sharing grants for every tool you no longer use.

4. Two zero-days in the tools that manage other machines

Researcher Sean Bales published exploits for N-able N-central, dubbed 4EverAble2Shell and Agent2Shell, on 12 November. That was two days after N-able shipped emergency fixes for earlier bugs. Attackers began probing within hours. Researchers counted at least 21 compromised systems, 19 of them on the latest security update, being used to mine cryptocurrency.

On Windows, a public proof of concept called SignDrop claims a local privilege escalation to SYSTEM through the CLFS driver. Microsoft has not assigned a CVE or shipped a fix, and no attacks have been reported so far.

If an MSP or internal IT team runs remote management tools across your estate, ask them this week how N-central is exposed to the internet.

5. GitHub: pull_request_target goes off by default on 2 December, and six Copilot models retire on 14 December

GitHub announced on 2 November that the `pull_request_target` trigger will be blocked by default in public repositories from 2 December. Two two-hour brownouts ran on 9 and 16 November, and blocked workflows failed with an error annotation. This trigger runs with the base repository's secrets and has been behind many CI compromises. If your release automation relies on it, you need to move to an explicit event policy or redesign the workflow.

Separately, a GitHub changelog post on 13 November schedules six models for retirement from Copilot chat, inline edits and completions on 14 December, including Gemini 3.8 Flash and Grok 4.6. Teams that standardised on a particular model should test its replacement before then. Our LLMOps work treats model retirement as a routine release event for this reason.

6. Anthropic's IPO prices this week

Anthropic will price its IPO on Wednesday 18 November and begin trading on Nasdaq under the ticker ANT on Thursday 19 November. It is offering 825 million shares at $70–78, which implies a valuation of between $1.3tn and $1.45tn (Yahoo Finance). At the top of the range, it would raise up to $64.35bn (CNBC).

For buyers, the useful change is disclosure. A listed model vendor publishes audited financials and risk factors every quarter, so you can read its finances directly in its filings instead of relying on press reports. It does not change your contract or your rate limits. Your exit plan should still cover the case where a model you depend on is repriced or retired.

7. Stripe closes its OpenRouter acquisition

Stripe said on 10 November that OpenRouter is now part of the company. OpenRouter is a model router that, at the time of the deal announcement in August, offered 400+ models from 80+ providers. By its own figures, more than 10 million developers and companies use it. OpenRouter says the product, name and integrations will not change, and it has reaffirmed its neutrality.

This is relevant if you route LLM traffic through OpenRouter. Your inference gateway and your payments provider are now the same company. That may be convenient for usage-based billing of AI features, but it adds concentration risk. It is worth recording in your vendor register.

8. EU Digital Omnibus: the GDPR rewrite moves through Parliament committee

The two halves of the Digital Omnibus are on different tracks. The AI half is already law as Regulation (EU) 2026/1744, adopted in June, and it postponed high-risk deadlines in the AI Act. The data half, covering GDPR, ePrivacy, NIS2 and DORA, is not law yet.

The Council agreed its mandate on 21 October. It keeps the current definition of personal data, sets a 96-hour notification deadline for high-risk breaches, and routes notifications through national entry points rather than a single EU portal (Technology.org). Parliament's IMCO committee voted on its opinion on 10 November. Parliament as a whole still has no position, so trilogue talks have not started.

Two things to watch. First, the Commission's expected decision naming AWS and Microsoft Azure as DMA gatekeepers. Bloomberg reported it for November, but the decision was still a draft and could slip. Second, if you build to GDPR today, you should not redesign anything yet. Nothing in the data Omnibus applies until it is adopted.

9. UK: Ofcom fines Meta £450,000 for inaccurate answers

On 12 November, Ofcom fined Meta £450,000 because its answers to statutory information requests about WhatsApp Business messaging were inaccurate. Ofcom found five contraventions across two notices. The fine includes a 25% discount for admitting liability and settling.

The amount is small for Meta, but the principle applies to everyone. UK regulators now treat inaccurate data in an information request as a breach in its own right. If you run a platform within Ofcom's scope, whoever answers regulators needs data you can reproduce, not estimates.

10. Mobile: iOS 27.2.1 security fixes and an Android age-range API

Apple released iOS and iPadOS 27.2.1 on 9 November, fixing nine vulnerabilities. It also shipped 26.7.3 and 18.7.12 for older devices.

On Android, Google Play services v26.44 (9 November) added an API that lets apps query the age range of the device user, and v26.45 added developer features for Wallet and payments. If your app has age-assurance duties, the platform-level signal is worth assessing before you build your own.

Deep analysis: the wfx-loot campaign and why your CI is now the soft target

What happened

From about 12:45 UTC on 10 November, an attacker used stolen credentials for 4,837 GitHub accounts against every repository those accounts could write to, according to Aikido Security, a security vendor that named the campaign wfx-loot. Each attack had three steps:

Row of brass padlocks on a fence with one hanging open
wfx-loot did not break in: it used accounts that already had write access.
  1. Probe. Open a pull request titled `docs: dependency audit notes` from a `wfx-docs` branch to confirm write access.
  2. Payload. Push a shell script of about 1,500 lines in one of three ways: injected into an existing workflow (`ci: refresh build cache`); as an npm `postinstall` hook (`chore: add ci helper`); or as a new workflow on a throwaway `chore/` branch.
  3. Clean up. Revert the change quickly and delete the workflow run and its logs.

Because the run is deleted, a missing run number is often the only trace. Aikido says this is a more reliable sign of compromise than looking for artefacts. If the stealer ran in a repository, you should assume every secret and cloud role available to its workflows has been taken.

The figures below come from security vendors analysing public GitHub data. They undercount private repositories, and the stealer had dedicated settings for private repositories.

The numbers

MeasureFigureSource
Compromised GitHub accounts used4,837Aikido
Repositories that received probe pull requests (to 11 Nov)61,764The Hacker News, citing Aikido
Probe pull requests by 12 Novover 66,200 (about 58,900 still open)OX Security
New probe pull requests on 12 Nov aloneover 4,100OX Security
Probes sent through collaborator or organisation access10,431, of which 92% were still openAikido
Repositories where the stealer ranat least 1,806Aikido
Public repositories with malicious commits on the default branch, 10 to 12 Nov8,093SafeDep, via The Hacker News
…delivered by workflow injection7,309same
…delivered by npm `postinstall` hook771same
Compromised users who publish packages (npm, Go, Packagist, PyPI, crates)at least 50Aikido
WebdriverIO repositories hit through its bot account28The Hacker News
The wfx-loot campaign in numbers

What the stealer takes, and what it leaves behind

Aikido's write-up describes four stages.

Collect. The stealer takes:

  • Actions secrets and variables, the `GITHUB_TOKEN`, and OIDC and runtime tokens;
  • credential files for AWS, Azure, GCP, Kubernetes, npm, PyPI, SSH, Vault and Terraform;
  • keys for AI services and AI coding tools;
  • shell histories, and every `.env`, `.pem` and `kubeconfig` file under the home directory;
  • secrets held in the memory of the runner's own processes.

On self-hosted runners, it also takes instance metadata credentials.

Escalate. It exchanges GitHub OIDC tokens for AWS, GCP and Azure sessions. On AWS, it creates long-lived access keys for IAM users.

Persist. This is the part that should worry engineering leaders. Three steps are designed to survive credential rotation:

  • It rewrites an AWS role's trust policy to `"Principal": {"AWS": "*"}`, which opens the role to any AWS account.
  • It adds federated identity credentials named `wfx-ci-…` to Azure AD applications. The compromised repository can then authenticate to Azure without a stored secret, even after every token has been rotated.
  • It posts forged `success` status checks under the names of the repository's required checks. Aikido explains that branch protection matches on the check's name, not on which app posted it, so the forged checks clear merge gates.

Exfiltrate. In this campaign, data went out by HTTP POST to two servers. SafeDep later found a third channel through Telegram, used alone in about a fifth of 200 sampled payload commits (The Hacker News).

Rotating secrets is necessary but not sufficient. If the stealer ran with cloud access, look for the trust-policy rewrite and the federated credential, or you will rotate keys while the back door stays open.

Context: this has been building since October

This is not a one-off. The accounts used this week overlap with October's GhostAction wave. Accounts hit in October were about 1,000 times more likely than other active accounts to be used this week, and in more than 100 repositories the new stealer was added to October's malicious workflow files (The Hacker News). Aikido does not attribute both waves to the same actor. Its point is that stolen credentials get resold, so an account compromised a month ago and never cleaned up becomes the next attacker's way in.

The defaults are improving, but slowly. npm 12 stopped running install scripts automatically from 8 July. Teams and CI images still on older npm versions remain exposed to the `postinstall` variant. GitHub's `pull_request_target` block on 2 December closes a different, older route. Neither change would have stopped wfx-loot, because the attacker was not an outsider sending a fork PR. It was an account with real write access.

Who is affected by wfx-loot

You are…Your exposure
A product company with private repos and GitHub Actions deploying to AWS, Azure or GCPHighest. Any developer or contractor account with write access is a potential way in, and OIDC deploy roles are a direct path to production
An open-source maintainer or package publisherHigh. Publishing triggers carried the stealer into at least one Packagist package and two Go modules
A company that uses self-hosted runnersHigher than with hosted runners, because the stealer also takes instance metadata credentials and Docker state
A buyer whose software is built by an agency or freelancersIndirect but real. Their accounts can write to your repositories, and 10,431 of the probes came through collaborator or organisation access
A team with no CI cloud credentials (manual deploys)Lower. Developer laptops are still exposed through the npm hook
Exposure to wfx-loot by type of organisation

What it means for companies building or running software

1. CI is now production. A workflow that can assume a deploy role has the same blast radius as an admin console login. Most teams review application code carefully and then merge workflow YAML with a glance. Changes to workflows should get the same review as changes to IAM policies, because in effect that is what they are.

Illustrated stone canal lock narrowing water into a single controlled stream
Treat CI like production: narrow trust so access flows only where it should.

2. Identity hygiene is the real perimeter. The attacker did not exploit a GitHub bug. They signed in. Phishing-resistant authentication, short-lived tokens and fast offboarding of departed contributors matter more than any scanner. Our passkeys guide covers the rollout side.

3. OIDC is better than static keys, but only if the trust is narrow. The stealer minted OIDC tokens and exchanged them for cloud sessions. That works against any role whose trust policy accepts tokens from an entire organisation or from any branch. Tie each role to one repository and one protected environment or branch.

4. "Required checks passed" is not proof. If your merge gates accept a status from any source, a stolen token can satisfy them. GitHub lets you set the expected source of a required check, such as a specific GitHub App. Use it.

5. Not everything needs rebuilding. For most small teams, triage takes a day. Hardening is a short, bounded piece of work: usually one to two weeks of part-time effort for a senior engineer across a handful of repositories. It does not need a platform rebuild. Do not start a migration away from GitHub over this. The weaknesses are in configuration and identity, and they would follow you to any other platform.

A practical checklist

Today: find out whether you were hit (hours, not days)

  • Search every repository you can write to, from 10 November onwards, for: branches `wfx-docs`, `chore/` followed by eight random characters, or `ci-check-` followed by four digits; pull requests titled `docs: dependency audit notes`; and commits titled `ci: refresh build cache`, `chore: add ci helper`, `chore: add dependency audit workflow`, `ci: restore` or `ci: cleanup` (Aikido's indicators).
  • Look for gaps in workflow run numbers around those commits.
  • Check outbound logs for connections to the published exfiltration servers.
  • Block Packagist `digitalelvis/neuronai-studio` 3.14.3 and Go `github.com/onuragtas/docker-env` v1.0.278 and v1.0.280.

If you find anything: contain (same day)

  • Revoke the affected account's tokens, OAuth app grants, SSH keys and sessions, and enforce 2FA. SafeDep's advice, via The Hacker News, is that any account that opened a probe PR is compromised, even if the stealer never ran.
  • Rotate every secret and variable the repository could read, including AI provider keys.
  • Treat any machine that ran `npm install` on an affected commit as compromised.
  • Hunt for persistence. Entra ID: federated credentials named `wfx-ci-` that you did not create. AWS CloudTrail:* sessions named `wfx-`, followed by `CreateAccessKey` or `UpdateAssumeRolePolicy`; also check every role trust policy for `"AWS": ""`. GCP: `GenerateAccessToken` and `SignJwt` calls from your GitHub workload identity pool. Vault: JWT logins from the GitHub OIDC issuer.

This month: harden (one to two weeks of focused effort)

  • Set the organisation default `GITHUB_TOKEN` permission to read-only and grant write access per job.
  • Put deploy credentials behind GitHub environments with required reviewers. A push to a branch should never reach production secrets on its own.
  • Narrow every OIDC trust policy to a specific repository and environment.
  • Set an expected source on required status checks.
  • Require review from CODEOWNERS for `.github/workflows/**` and `package.json` changes.
  • Move to fine-grained, expiring tokens. Enforce SSO and phishing-resistant MFA for everyone with write access, including contractors.
  • Use ephemeral self-hosted runners with restricted outbound network access, if you use self-hosted runners at all.
  • Audit collaborators and remove anyone who no longer needs write access. Tie offboarding to your contract end dates.
  • Audit every `pull_request_target` workflow before 2 December.

This quarter: make it routine

  • Add CI configuration to your regular software audit and threat model.
  • Alert on workflow file changes and on new federated credentials or trust-policy edits in your cloud accounts.
  • Rehearse a full rotation of secrets for one repository, so you know how long it really takes.

If you would like a second pair of eyes, this is the kind of work our DevOps and cyber security engagements cover. A senior engineer reviews your pipelines, cloud trust relationships and access model, and you keep everything that is changed. For ongoing patching across a stack like the one in this week's Patch Tuesday, see our support and evolution model. You can talk to the founders directly: conversations start online and go straight to the engineers who would do the work.

Частые вопросы

How do I know if the wfx-loot stealer ran in my repository?

Look for the branch, pull request and commit names listed in the checklist above, from 10 November onwards. Then check for missing workflow run numbers around those commits. The attacker deleted runs and logs, so the absence of artefacts proves nothing. Aikido says a missing run number is the most reliable sign.

We only use private repositories. Are we safe?

Not necessarily. The published counts cover public repositories, because those are the ones researchers can see. The stealer had dedicated settings for private repositories, and it works through any account with write access. Run the same checks on private repositories.

Is rotating our secrets enough?

No, not if the stealer ran with cloud access. It is designed to survive rotation in two ways: by rewriting AWS role trust policies to accept any account, and by adding Azure federated credentials. Remove those first, then rotate.

Should we stop using OIDC and go back to static cloud keys?

No. Short-lived OIDC credentials are still safer than long-lived keys. The fix is to narrow the trust, so that each role accepts tokens only from one repository and one protected environment. Add human approval before deploy credentials are issued.

With more than 1,000 Microsoft CVEs in one month, what should a small company patch first?

Start with the exploited Outlook flaw CVE-2026-124194. Then fix internet-facing servers: Exchange, Schannel, SharePoint and IIS. Then developer machines, including VS Code. Then everything else through automatic update rings. Also plan to move off .NET 8 and PowerShell 7.4, which have left support.

Does Anthropic's IPO change anything for teams building on Claude?

Not your contract or your API. What changes is that more information about the vendor will be public. Whichever model vendor you use, keep an abstraction layer and an evaluation set, so you can switch models if pricing or availability changes.

Do we need to act on the EU Digital Omnibus now?

On the data side, no. The GDPR, ePrivacy and NIS2 changes are still under negotiation, and Parliament has no position yet. The AI Omnibus is already law. Separately, the AI Act transparency obligations take effect on 2 December, so check those now with our guide.

Pranas MickeviciusChief Executive Officer
Опубликовано:
17 ноября 2026 г.
  • GitHub Actions
  • Supply chain security
  • CI/CD
  • Patch Tuesday
  • AWS
  • Cloud identity
  • Weekly briefing

Заметки о разработке раз в месяц

Чему мы учимся, создавая SaaS-платформы, приложения и функции на базе ИИ: коротко, по делу и от основателей, которые сами делают эту работу.

Одно письмо в месяц. Отписаться можно в любой момент. Политика конфиденциальности

Похожие материалы

Больше материалов

Расскажите, что вы создаёте. Основатель ответит в течение одного рабочего дня и сам будет разрабатывать проект вместе с вами.

Мы внесём ваши данные в нашу CRM, чтобы связаться с вами по вашему запросу. Подробнее — в нашей политике конфиденциальности.