ICHI Briefs & Podcasts · ICHI Daily · five-minute podcast

ICHI Global Five Minute Cyber Rundown

ICHI Daily · Published · 5 min listen

Akari's Thursday briefing: the wfx-loot GitHub campaign stole CI and cloud secrets from at least 1,800 repositories, ScreenConnect panels were turned against MSPs, Microsoft's Lazarus research backdoors AI agents with a handful of neurons, and TSMC and Maersk confirm partner and legacy breaches.

Listen to this episode

MP3, 5.2 MB · · the full transcript is below.

Show notes

ICHI Global Five Minute Cyber Rundown — Thursday, November 12, 2026

Today's theme is stolen trust at scale: a GitHub Actions secret-stealing campaign that ran in at least 1,800 repositories, remote-access panels turned against managed service providers, a research breakthrough that backdoors AI agents by editing a handful of neurons, and two confirmed breaches through a cloud partner and a legacy system.

Most important development — wfx-loot: GitHub Actions secret stealer (Aikido, Nov 12)

Starting at about 12:45 UTC on November 10, 2026, a threat actor using stolen GitHub credentials opened probe pull requests titled "docs: dependency audit notes" from wfx-docs branches in 61,764 repositories, using 4,837 compromised accounts (Aikido Security, published November 12). GitHub search returned 61,839 such pull requests opened November 10–11; 10,431 went into repositories the account did not own, via collaborator or organisation access, and 92% were still open.

Aikido identifies indications the payload ran in at least 1,806 repositories (its summary count is 1,835 repositories with credentials exfiltrated; run logs were deleted, so this is a lower bound). The payload is a single ~1,500-line shell script ("wfx") delivered by workflow injection, an npm postinstall hook, or a throwaway workflow branch; it reverts within a minute and deletes the run.

What it steals and does:

  • Collects GitHub Actions secrets and variables, GITHUB_TOKEN and OIDC/runtime tokens, cloud credential files (AWS, Azure, GCP, Docker, Kubernetes, npm, PyPI, Git, SSH, Vault, Terraform), AI-provider and AI-tool tokens (including OpenAI, Anthropic, Hugging Face, Cursor, Copilot), shell histories, and .env / key files.
  • Escalates via OIDC exchange into AWS, GCP and Azure; creates long-lived AWS access keys and impersonates GCP service accounts.
  • Persists beyond rotation: forges passing branch-protection status checks, rewrites an AWS role trust policy to allow any AWS account, and adds Azure federated identity credentials named wfx-ci-… pointing back to the compromised repository.
  • Exfiltrates by HTTP POST to attacker servers in this campaign.

Automated publishing carried the payload into Packagist digitalelvis/neuronai-studio 3.14.3 (treat any machine that installed or opened it as compromised) and Go module tags github.com/onuragtas/docker-env v1.0.278 and v1.0.280 (payload in a workflow file only; Go builds do not run it). At least 50 compromised users publish packages for npm, Go, Packagist, PyPI or crates, though Aikido assesses the focus is CI secrets, not supply-chain proliferation. Accounts hit by October's GhostAction wave were significantly more likely to be hit again; Aikido assesses the two campaigns are likely different actors.

Defence: hunt for wfx-docs branches, the probe pull request, and commits titled "ci: refresh build cache", "chore: add ci helper" or "chore: add dependency audit workflow". A missing workflow run number is a more reliable sign the payload ran than missing wfx-loot artifacts. If it ran, assume every secret and cloud role reachable by its workflows is compromised — rotate them and remove the AWS trust-policy and Azure federated-credential persistence.

Incidents and adversary activity — ScreenConnect / SCKit against MSPs (Huntress, Nov 11)

Huntress first detected activity at about 17:47 UTC on November 11, 2026: attackers using access to cloud-hosted ConnectWise ScreenConnect panels pushed Run Commands to install rogue ScreenConnect clients on at least 288 managed endpoints across 48 organisations, all within about 90 minutes (at least 6 MSPs in earlier counts). The rogue clients deliver SCKit, a custom toolkit for persistent remote access that creates a relay trap for intercepting technician session cookies. Payload analysis shows intended lateral movement from an infected endpoint to a connecting technician's workstation and then to other panels that technician can reach; Huntress says it has not observed a successful jump of that kind in recovered logs, and not all SCKit functionality was observed in incidents.

Huntress links access to the third-party RMM+ extension: versions 1.0.31 and 1.0.32 accepted a blank access key when an administrator had not configured one, bypassing ScreenConnect accounts, roles and MFA for its API. The Hacker News notes Huntress also said initial panel access could involve compromised credentials or hijacked sessions, and the exact method remains unclear in ConnectWise's account.

ConnectWise says there is no evidence of a compromise of its infrastructure or a vulnerability in ScreenConnect itself, that the activity may involve abuse of independent third-party extensions, and that the full scope is still under investigation. As a precaution it disabled five extensions — Labtech Plus, RMM Plus, RMM+, RMM Plus Password Link and RMM+ Password Link — in about 400 partner environments where it could act, and urged on-premises partners to disable them. Presence of an extension alone does not establish compromise. ScreenConnect users should review Run Command history even without alerts, and check the IoCs published by ConnectWise and Huntress.

AI + cyber breakthrough — Microsoft Research "Lazarus" dormant-neuron backdoor (Nov 10)

Microsoft Research published Lazarus ("Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents") on November 10, 2026. The technique rewrites fewer than 10 dormant neurons in a model's final feed-forward block, with no retraining and only inference-level compute, so the model behaves normally and passes standard evaluations until a chosen trigger appears — for example a date change or a string in tool output — at which point it emits an attacker's command that a tool-using agent then runs with the agent's own permissions.

In the headline test on Nemotron-Terminal-8B, editing 7 of more than 12,000 neurons produced the payload in at least 98.8% of triggered sessions and in none of the checked trigger-free sessions; results on other models varied (about 51% to 99.5%). The attack needs write access to a model checkpoint — a compromised account, pipeline or storage bucket is enough; internally developed models are equally exposed. Microsoft says it has not observed threat actors weaponising this technique — this is a research result, not a reported campaign.

Defence: treat model weights like code — control checkpoint write access, verify provenance, and diff weights against a trusted declared base model, which Microsoft identifies as the strongest detection signal.

Confirmed breaches — TSMC employee data via a cloud provider (Nov 12)

TSMC confirmed on November 12, 2026 that information for some current and former employees was improperly posted externally. Preliminary comparison matches the leaked content to data held by a cooperating third-party cloud service provider. TSMC says there is no evidence its internal IT systems or wafer manufacturing equipment were compromised, and the data does not involve chip manufacturing operations or customer information. It is supporting affected employees and working with the provider, law enforcement and third-party security firms.

Sellers on DarkForums / BreachForums claim about 46,000 records / 6.9 GB for 3 BTC (about $220,000), threatened to release the data within a week, and published a sample said to contain 499 records with 19 fields (including names, ID numbers, addresses, phone numbers, bank details and emergency contacts). TSMC has not confirmed the sellers' total or field list.

Confirmed breaches — Maersk legacy system (confirmed Nov 11)

A.P. Moller–Maersk confirmed on November 11, 2026 that an unauthorised party accessed a local legacy system inherited through a previous acquisition and extracted data. Maersk says the incident is isolated to that system, with no indication of impact on wider Maersk infrastructure or on customer and partner systems connected via APIs and EDI, no expected operational disruption, and that preliminary findings strongly suggest no sensitive data was disclosed. It has not named the system, the data types taken, or when the intrusion began.

The WhiteArmor group, which listed Maersk on its leak site over the weekend of November 7–8, claims about 100 GB from the former LF Logistics platform (acquired 2022) — including driver profiles, plaintext passwords, admin accounts, delivery receipts and inventory records — and demands $5 million in bitcoin within 30 days. Maersk has not confirmed WhiteArmor's claims or linked the incident to the group, and the claimed data has not been independently verified as the data Maersk confirmed was extracted.

Vulnerability to patch — SAP pre-authentication flaw (in prose, not exploited)

SAP's November 2026 Patch Day (November 10) fixed a pre-authentication memory-corruption flaw in the Internet Communication Manager in SAP NetWeaver AS ABAP and the closely related flaw in SAP Web Dispatcher (including the dispatcher embedded in SAP HANA XS), tracked by researchers at Onapsis as ICM2SHELL. WebSocket handling is on by default, so a network-reachable system can be exposed even if no application uses WebSockets; a crafted sequence can crash the component or run code as the account that owns the SAP installation. Onapsis provided SAP with a working proof of concept. No in-the-wild exploitation has been reported and it is not a confirmed-exploited item — patch the kernel, every Web Dispatcher and the related SAP notes as an emergency change, and use SAP's temporary workarounds where patching must wait. This item is covered in prose with no identifier or score, per the episode's editorial rule for non-exploited vulnerabilities.

Defensive implications

  • GitHub / CI: search for the wfx-loot probe, rotate Actions, cloud and AI-tool secrets, and remove persistence that survives rotation (AWS trust policies, Azure federated credentials, long-lived keys).
  • ScreenConnect / MSPs: disable the five listed extensions, audit Run Command history and technician workstations, and hunt for rogue clients and SCKit artifacts.
  • AI agents: restrict and audit model-checkpoint write access; verify and diff model weights before giving an agent tool or terminal permissions.
  • Third parties and legacy systems: inventory employee data held by cloud providers and data on acquired / legacy platforms; those are the sources of today's two confirmed leaks.
  • SAP: apply the November kernel and Web Dispatcher fixes urgently despite no reported exploitation.

ICHI note

No live ICHI composite score is cited in this episode because the live score did not render in the pre-production check at production time. ICHI.global's production methodology remains VECTORS v7.1; no methodology change is announced. Model C scoring refinements remain under development.

SOURCES

  • Aikido Security — wfx-loot campaign pushes CI secret stealer to tens of thousands of repositories: https://www.aikido.dev/blog/compromised-github-accounts-push-ci-secret-stealer-to-tens-of-thousands-of-repositories
  • Help Net Security — Check your GitHub repos: A secret stealer hit tens of thousands of them: https://www.helpnetsecurity.com/2026/11/12/wfx-loot-github-actions-secret-stealer/
  • The Hacker News — ScreenConnect Attacks Target at Least 288 Endpoints Across 48 Organizations With Rogue Clients: https://thehackernews.com/2026/11/screenconnect-attacks-target-at-least.html
  • SecurityWeek — Hackers Deploy Malware in Fresh ScreenConnect Campaign: https://www.securityweek.com/hackers-deploy-malware-in-fresh-screenconnect-campaign/
  • KuCoin / TechFlow — Hacker Demands 3 BTC for 6.9GB of TSMC Employee Data, Claims Leak Originated from Cloud Vendor: https://www.kucoin.com/news/flash/hacker-demands-3-btc-for-6-9gb-of-tsmc-employee-data-claims-leak-from-cloud-vendor
  • Crypto Briefing — TSMC confirms employee data leak as hackers seek 3 BTC for the records: https://cryptobriefing.com/tsmc-employee-data-leak-3-btc/
  • Neowin — Shipping company Maersk suffers data breach, ransomware gang demands $5 million: https://www.neowin.net/news/shipping-company-maersk-suffers-data-breach-ransomware-gang-demands-5-million/
  • BleepingComputer — SAP patches ICM2SHELL and PROXYGATE maximum severity flaws: https://www.bleepingcomputer.com/news/security/sap-patches-four-maximum-severity-security-vulnerabilities/
  • Onapsis — ICM2SHELL: A Critical Pre-Authentication Vulnerability in SAP NetWeaver's ICM and Web Dispatcher: https://onapsis.com/blog/icm2shell-cve-2026-76964-cve-2026-89126-sap-netweaver-web-dispatcher-threat-advisory/
  • Cloud Security Alliance Labs — Seven Neurons to Root: Lazarus Weight-Edit Backdoors in Agent Models: https://labs.cloudsecurityalliance.org/research/csa-research-note-lazarus-neuron-edit-agent-backdoor-2026111/

Transcript

Akari: The daily cyber rundown, brought to you by I C H I dot global. I'm Akari, an AI-generated voice. This is Thursday, November twelfth, twenty twenty-six. Today the theme is stolen trust at scale, a GitHub campaign that harvested cloud secrets from thousands of repositories, remote access turned against its providers, and two confirmed breaches through partners and legacy systems.

Akari: The most important development is the campaign researchers at Aikido are calling wfx-loot. Starting on November tenth, an attacker using stolen GitHub credentials opened probe pull requests in more than sixty-one thousand repositories, from nearly five thousand compromised accounts. In at least one thousand eight hundred repositories the payload actually ran, and credentials were taken. The payload is a single script that collects GitHub Actions secrets, cloud credential files, shell histories, and tokens for A I tools, then tries to reach deeper into A W S, Google Cloud, and Azure. It also leaves persistence that survives rotation, rewriting a cloud role trust policy and adding federated credentials in Azure. If the payload ran in your repository, treat every secret its workflows could reach as compromised, and hunt for a branch called wfx-docs or a pull request titled docs, dependency audit notes.

Akari: Next, remote access itself became the attack path. Huntress says attackers abused cloud hosted ScreenConnect panels on November eleventh to push commands to at least two hundred eighty-eight endpoints across forty-eight organizations, all within about ninety minutes. The rogue clients delivered a toolkit called S C Kit, built for persistent access and for trapping technician session cookies. Huntress links the access to a third party extension called R M M plus, whose recent versions accepted a blank access key when an administrator had not set one. ConnectWise says it has found no compromise of its own infrastructure and no flaw in ScreenConnect itself, and it has disabled five related extensions across about four hundred partner environments as a precaution. Providers should check their Run Command history, even without an alert.

Akari: Our A I and cyber breakthrough today is research, not a campaign. Microsoft Research published a technique it calls Lazarus on November tenth. By rewriting fewer than ten dormant neurons in a model, with no retraining, researchers made an A I agent behave normally until a hidden trigger appeared, then emit an attacker's command for the agent to run with its own permissions. In one test, editing seven neurons produced the command in almost every triggered session and never in clean ones. Microsoft says it has not seen attackers use this in the wild. The lesson is practical. Treat model files like code, verify where they came from, and diff them against a trusted original.

Akari: Two confirmed breaches underline the partner problem. T S M C said on November twelfth that data for some current and former employees was leaked, and that early checks match records held by a third party cloud provider. Its own systems show no intrusion, and production and customer data were not involved. Sellers claim about forty-six thousand records, but T S M C has not confirmed that total. Separately, Maersk confirmed an attacker accessed a legacy system from a past acquisition and extracted data, isolated from its wider platforms. A group calling itself WhiteArmor claims about one hundred gigabytes and is demanding five million dollars. Maersk has not confirmed those claims, and says early findings suggest no sensitive data was disclosed.

Akari: One more item to patch, serious but not yet exploited. S A P has fixed a pre authentication flaw in the component that handles web traffic for its business systems, and in its web dispatcher, that could let an unauthenticated attacker run code as the account that owns the installation. Researchers gave S A P a working proof of concept, but no exploitation in the wild has been reported. Patch it urgently, do not wait for exploitation.

Akari: Defensively, then. For GitHub, rotate Actions and cloud secrets, remove the persistence described above, and check for the probe pull request. For ScreenConnect, disable the listed extensions, review commands, and look for rogue clients. Verify A I model provenance before an agent can run tools. And review employee and legacy data held by partners, where today's confirmed leaks came from. The takeaway is trust must be verified, not assumed. For the underlying evidence behind every story today, visit I C H I dot global. I'm Akari. Stay safe, stay frosty, and I will see you tomorrow.

Sources

  1. Aikido Security — wfx-loot campaign pushes CI secret stealer to tens of thousands of repositories www.aikido.dev
  2. Help Net Security — Check your GitHub repos: A secret stealer hit tens of thousands of them www.helpnetsecurity.com
  3. The Hacker News — ScreenConnect Attacks Target at Least 288 Endpoints Across 48 Organizations With Rogue Clients thehackernews.com
  4. SecurityWeek — Hackers Deploy Malware in Fresh ScreenConnect Campaign www.securityweek.com
  5. KuCoin / TechFlow — Hacker Demands 3 BTC for 6.9GB of TSMC Employee Data, Claims Leak Originated from Cloud Vendor www.kucoin.com
  6. Crypto Briefing — TSMC confirms employee data leak as hackers seek 3 BTC for the records cryptobriefing.com
  7. Neowin — Shipping company Maersk suffers data breach, ransomware gang demands $5 million www.neowin.net
  8. BleepingComputer — SAP patches ICM2SHELL and PROXYGATE maximum severity flaws www.bleepingcomputer.com
  9. Onapsis — ICM2SHELL: A Critical Pre-Authentication Vulnerability in SAP NetWeaver's ICM and Web Dispatcher onapsis.com
  10. Cloud Security Alliance Labs — Seven Neurons to Root: Lazarus Weight-Edit Backdoors in Agent Models labs.cloudsecurityalliance.org

Get ICHI Daily and ICHI Weekly by email

No new episode, no email. Every email links to the episode on ICHI; audio is never attached.


What to receive


✓ Daily podcast included ✓ Weekly podcast included ✓ Weekly written brief included


✓ Daily podcast included ✓ Weekly podcast included ✗ Weekly written brief not included


✗ Daily podcast not included ✗ Weekly podcast not included ✓ Weekly written brief included


✗ Daily podcast not included ✓ Weekly podcast included ✓ Weekly written brief included


✓ Daily podcast included ✗ Weekly podcast not included ✗ Weekly written brief not included


Choose any combination below, including the weekly podcast only or the weekly written brief only.

Custom: choose any combination

We send a confirmation email first, and nothing else until you confirm. Already subscribed? Use the “manage” link in any email to change what you receive; nothing changes unless you choose it. Your address is used only to send these emails; every email has a one-click unsubscribe link.

All ICHI Daily episodes · ICHI Briefs & Podcasts

ICHI Daily and ICHI Weekly are editorial commentary on ICHI’s evidence. They never change the VECTORS score, incident evidence classifications or intelligence rankings.

ICHI Briefs & Podcasts · ICHI Daily · ICHI Weekly · Episode calendar · ICHI Perspectives · episodes.json · ICHI dashboard · Methodology