Cyber Incidents 17-18 Nov 2026: What Is Confirmed
TL;DR · the short version
A threat-intel digest for 17-18 November 2026 listed 21 items. We checked each one. Most are real and sourced: hijacked Optimizely Web snippets that swapped crypto wallet addresses, skimmed cards and tried to deliver iOS exploit kits on 178 organisations' sites (Report URI); a Bitdefender-reported campaign that uses AGENTS.md, CLAUDE.md, session-start hooks and MCP entries to make AI coding agents run a credential stealer; ScreenConnect CVE-2026-84869 used to install a rogue ScreenConnect client on an MSP technician's PC (Huntress); six unpatched pre-login flaws in Danfoss AK-SM800A refrigeration controllers (Midnight Blue); Infoblox confirming unauthorized access through a compromised data migration vendor; Uber notifying people after attackers used stolen Fireworks AI credentials; a GitLab release with 37 security fixes; and an extortion attack that closed the Greiz emergency department for about two days. Some items are vendor research or forecasts, not incidents, and one digest label (Storm-2372 and EvilTokens on the AI coding agent campaign) is not supported by the source. Facts are as of 18 November 2026 and may change.
A threat-intel platform digest for 17 and 18 November 2026 listed 21 clusters. We treated each title as a lead and checked it against the vendor's research, the company's own statement, a CERT or regulator, the US Department of Justice or a reputable outlet. Most items are real and sourced, the biggest being a hijack of Optimizely Web snippets on 178 organisations' sites, a campaign that turns AI coding agent configuration files into a credential stealer, live exploitation of ScreenConnect CVE-2026-84869 against MSP technicians, unpatched Danfoss refrigeration controller flaws, and vendor-route breaches at Infoblox and Uber. Several are research or forecasts, not incidents, and one digest label (Storm-2372 and "EvilTokens" on the AI coding agent campaign) is not supported by the source.
Who this is for: security leaders, MSPs, platform and developer-tooling owners, OT and facilities owners, and anyone asked "are we exposed?" this week.
Independent analysis. PlayCISO is not affiliated with any organisation named. Facts are as reported on 18 November 2026 and may change. Five items have their own posts and are kept short here.
Confirmed vs not yet known
Confirmed by the company, a CERT, the DoJ or the researchers' own published data: malicious code in Optimizely Web snippets (Optimizely notice); 37 GitLab security fixes; Infoblox's unauthorized access via a vendor; Uber data accessed through Fireworks AI credentials; the extortion attack on the Greiz-Ronneburg hospital (district statement to MDR); a 43-year-old Kazakh national extradited from Italy in the 8-Digits Team case (charges, not a conviction); Bitly's statement that old links were repurposed; CERT/CC's macOS Unicode advisory.
Vendor research (one company's findings, not independently confirmed): Bitdefender, Huntress, Midnight Blue, Unit 42, SOCRadar, Kaspersky, Google GTIG, Darktrace, Lookout, Stormshield, NewCore.
Not published or not supported: any confirmed iPhone compromise from the Optimizely injection; any actor behind the AI coding agent campaign (the Storm-2372/EvilTokens tag); a CVE or patch for the Danfoss flaws; how many Infoblox customers are affected; whether Greiz patient data left the network.
The 21 items at a glance
| Item | What happened | Status | Who reported | More |
|---|---|---|---|---|
| Optimizely snippets | Injected JS swapped wallet addresses, skimmed cards, tried iOS exploit kits on 178 orgs' sites | Confirmed (injection); no device compromise confirmed | Optimizely; Report URI; Cynet; StepSecurity | below |
| PureLogs / PureRAT loader | Shipment ZIP, DLL side-loading, Python disguised as a PDF | Vendor research; family label is infrastructure-based | Unit 42 (via press) | below |
| AI coding agent config malware | AGENTS.md / CLAUDE.md, hooks and MCP entries make agents run a stealer | Vendor research; actor tag unsupported | Bitdefender Labs | post |
| M365 outage phishing | "Outage" lures after the 12 Nov outage | Confirmed by Microsoft | Microsoft Threat Intelligence | post |
| XaocMail / Phantom53 | ClickFix campaign, implant with DNS-over-HTTPS C2 | Vendor research | SOCRadar | below |
| Danfoss AK-SM800A | Six unpatched pre-login flaws | Research; CERT advisories; no exploitation reported | Midnight Blue; CCB; NCSC-IE; BSI | post |
| AI agents and LLMjacking | Agent frameworks in intrusions; stolen LLM access resold | Vendor research | Google GTIG | below |
| GraphDoor | Backdoor using SharePoint via Microsoft Graph for C2 | Vendor research | Kaspersky | below |
| ScreenConnect CVE-2026-84869 | Rogue ScreenConnect client installed on an MSP technician PC | Exploited (CISA KEV); one compromise confirmed | Huntress | post |
| Claude memory stores as C2 | Lab proof of concept | Research, not seen in the wild | Darktrace | below |
| GravityRAT job lures | Fake recruitment sites vs Indian defence job seekers | Vendor research; date disputed | Lookout (via press) | below |
| Trigram | Fake job, conference and developer lures, cross-platform stealer | Vendor research; not fully verified | Stormshield (via press) | below |
| macOS Unicode collisions | ß vs ss path tricks bypass exclusions, overwrite files | CERT advisory, CVEs assigned | CERT/CC VU#604762 | below |
| 2027 outlook | AI, credential theft, ransomware forecasts | Forecasts, not incidents | SpyCloud; ZeroFox | below |
| GitLab | 37 security fixes incl. CI/CD, Pages, MCP server | Confirmed | GitLab | below |
| Grok Bot check-ins | 31 unattended data requests, all blocked | Vendor research; xAI response quoted | NewCore | below |
| Old Bitly links | 2009-2014 links in X posts repurposed | Confirmed by Bitly | CNN; Bitly | below |
| 8-Digits Team | Alleged member extradited from Italy | Charged, not convicted | US DoJ | below |
| Uber / Fireworks AI | Stolen vendor credentials used to reach Uber data | Confirmed by Uber | Uber notices; TechCrunch | post |
| Greiz-Ronneburg hospital | Extortion attack, emergency department closed about 2 days | Confirmed by the district | MDR; Landkreis Greiz | below |
| Infoblox | Company information accessed via a migration vendor | Confirmed; data scope unconfirmed | Infoblox via TechCrunch | post |
Supply chain and third parties
Optimizely Web snippets hijacked
Optimizely's incident notice says an attacker "exploited a vulnerability in Experimentation" and added malicious JavaScript to some customer projects; it says the flaw is fixed and the code removed. Report URI checked 1,611 snippets and found the code in 188 belonging to 178 organisations, among them CNN, BBC, Zoom, HSBC and IKEA, per The Hacker News. The first injection was at 04:21 UTC on 16 November; the payload swapped crypto wallet addresses, a card skimmer was added at 13:43 UTC, and on iPhone and iPad user agents a script tried to load the Coruna and DarkSword exploit kits. Report URI counts about $7,040 reaching attacker wallets (StepSecurity: about $6,600). No iPhone compromise has been confirmed: Cynet calls it "conditional delivery", not a successful compromise of a device.
Infoblox: access through a data migration vendor
Infoblox told TechCrunch (16 November) it is "investigating a security event involving unauthorized access to certain company information" through a compromised third-party data migration vendor, and has "no indication" its products, cloud services or software updates were involved. That the vendor did SharePoint migration work and that customer data was taken comes from unnamed sources, not Infoblox. Full write-up: Infoblox breach via a SharePoint migration vendor.
Uber and Fireworks AI
Uber's breach notice says attackers used credentials stolen from its AI vendor Fireworks AI to access Uber data held in the vendor's environment in October, including driver's licence details, and that Uber "immediately severed data connections" with the provider. Counts differ by filing; Security Affairs cites an Oregon filing of 149,346 people. Uber says it has seen no misuse. Full write-up: Uber data taken through stolen Fireworks AI credentials.
Old Bitly links in political, government and corporate X posts
CNN reported on 15 November that shortened links in old X posts from Donald Trump, the FBI, the US Department of Defense, Burger King, McDonald's and others now lead to unrelated or offensive sites. Bitly said it "identified an issue that allowed bad actors to repurpose certain old, inactive links" and suspended the accounts involved. The affected posts date from 2009 to 2014. Nothing published shows the X accounts themselves were compromised; Bitly has not explained the exact flaw.
Developers and AI agents
AI coding agent configurations turned into a credential stealer
Bitdefender Labs (17 November) found 22 malicious GitHub links whose AGENTS.md and CLAUDE.md files tell the assistant to run scripts/setup.js without showing it, and whose .claude/settings.json session-start hooks can run it once the developer trusts the folder. The payload steals source-control tokens, CI/CD secrets, cloud, container and package-registry credentials and browser data, checks which tokens work, then plants instructions, hooks and MCP entries for 14 agents and, with valid tokens, can modify workflows and try to republish packages. Bitdefender names no threat actor. The digest's Storm-2372 and "EvilTokens" tag (both names used for device-code phishing) is not in the report and we found no source linking them. Full write-up: AI coding agent config malware.
Darktrace: Claude memory stores as a lab C2 channel
Darktrace's 12 November post describes a proof of concept in which Claude Code, launched on an already compromised developer machine with the user's Claude Console API key, read jobs from and wrote results to Claude memory stores, and moved a 1 GB file out through the Files API. Darktrace says it is not a platform vulnerability, needs a prior compromise, and has not been seen in the wild. The defensive point: traffic to an approved AI API is not proof the activity is approved.
Grok Bot background check-ins
NewCore reported on 17 November that two Grok Bots on employee accounts made 31 unattended read requests over five days to Google Calendar, Gmail, Slack, GitHub and Notion; its MCP gateway held each for approval and none was approved. xAI support called them "built-in background check-ins". NewCore says xAI added an admin control on 12 November to turn the check-ins off, and that it is on by default. This is one vendor's research; NewCore reports no software vulnerability.
Trigram fake developer workflows
Stormshield, as reported by SecurityOnline and Daily Security Review, describes Trigram, a North Korea-aligned campaign using fake remote-engineering jobs, conference invitations and parcel pages to get developers to paste commands that install the cross-platform OverLord stealer. Stormshield links it with high confidence to the cluster Proofpoint tracks as UNK_DeadDrop. We did not open the Stormshield report or confirm its publication date; treat the details as reported.
Useful next, if this was relevant to you
GitLab fixes 37 vulnerabilities
GitLab's 16 November patch release (19.5.6, 19.4.8, 19.3.9) fixes 37 security issues: 4 high, 24 medium and 9 low. The high ones include an incorrect-authorization issue in CI/CD pipeline jobs (CVE-2026-18241); the list also has a path traversal in GitLab Pages (CVE-2026-11768) and an incorrect-authorization issue in the GitLab MCP server (CVE-2026-123142). GitLab.com and Dedicated are already patched; GitLab says self-managed installations should upgrade immediately.
macOS Unicode normalization collisions
CERT/CC note VU#604762 (16 November, revised 17 November) says tools that compare raw byte paths can be fooled on APFS, which treats different Unicode spellings such as ß and ss as the same file. A local attacker can bypass exclusion filters or overwrite files the process can write. Named CVEs cover HashiCorp go-slug (CVE-2026-14978), Deno (CVE-2026-49401), libarchive (CVE-2026-45110, fix pending) and setuptools (CVE-2026-59890, fixed in 83.0.0).
Remote access, phishing and malware
ScreenConnect CVE-2026-84869: rogue clients on technician PCs
Huntress (16 November) counted about 688,000 guest "run file" requests against 42 MSP ScreenConnect servers between 20 October and 16 November. On 12 November one technician's host client ran a pushed batch file that installed a rebranded ScreenConnect client within about nine seconds. Huntress confirms that one compromise and a test-file execution on a second workstation. The fix is host client 26.6.5 (released 8 September); updating the server alone does not fix it. Full write-up: CVE-2026-84869 rogue ScreenConnect clients.
Microsoft 365 outage-themed phishing
Microsoft Threat Intelligence published on 16 November a high-volume campaign seen on 13 November, using "Outage" subjects and claiming held mail would be deleted unless recipients clicked. Microsoft assessed it as likely credential phishing but did not confirm credential capture. The outage itself and what to tell staff are in our earlier post: Microsoft 365 outage, 12 Nov 2026.
XaocMail and Phantom53
SOCRadar (13 November) describes XaocMail, a self-hosted phishing platform run by a Russian-speaking actor, that sends ClickFix lures posing as Google Docs, Adobe Acrobat and Office 365. Victims install Phantom53, which hides command traffic in DNS-over-HTTPS queries to public resolvers and loads a Havoc agent. Targets include Brazil, Peru, India, Indonesia and Spain. SOCRadar says the code has "artifacts characteristic of development using Anthropic's Claude Code"; that is SOCRadar's assessment.
PureLogs and PureRAT via DLL side-loading
Unit 42 research reported on 16 November describes a shipment-themed ZIP with a renamed, signed Kingsoft PDF program that side-loads a malicious DLL, then runs a bundled Python interpreter on a file named like a PDF that actually holds encoded Python. The final payload talks to infrastructure Unit 42 associates with PureLogs and PureRAT; press coverage stresses that this link does not prove both families ran. We read secondary coverage, not the Unit 42 original.
GraphDoor uses SharePoint for C2
Kaspersky's "Operation ShadowCloud" research (17 November) describes GraphDoor, which authenticates to Microsoft as a malicious application and uses the Graph API to fetch commands from and stage data on SharePoint, and a sibling, GitDoor, that uses GitHub. Targets are government organisations in Asia and Africa; the backdoors were delivered through an abused MeshCentral agent. Kaspersky does not name the actor and says a Chinese-speaking group may be behind it, citing overlaps it says do not prove a link.
GravityRAT job lures
Lookout research, as reported by GBHackers, Cyberpress and others, describes fake recruitment portals (one called "Global Recruiter") using defence-company logos to get Indian defence job seekers to sideload GravityRAT, which collects contacts, messages, location and files. Lookout counted 14 detections in India over 12 months and links the activity to Pakistan-aligned actors it tracks as SpaceCobra and Cosmic Leopard. Outlets disagree on the disclosure date (5 or 17 November); the Lookout page was not reachable for us.
AI-assisted attacks and the outlook
AI agents in intrusions, and LLMjacking
Google's Threat Intelligence Group published on 16 November that threat actors use agentic frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI to "automate discrete phases and specific operational tasks during live intrusions", and that stolen LLM credentials and cloud resources are resold through underground proxy services. Coverage also describes a staging server that compared patched and unpatched code and drafted exploits without a human in the loop; secondary reports note many exploits did not work. Background: ARTEX and GTIG, exposed AI pentest consoles: what to hunt and checking for leaked LLM API keys.
2027 outlook reports
The digest did not say which report it meant. SpyCloud (17 November) and ZeroFox (18 November) both published 2027 forecasts: SpyCloud expects stolen sessions, tokens and vendor connections to do most of the damage and gives fully autonomous attacks low confidence; ZeroFox expects attacks on enterprise AI tools and the credentials they hold to grow and forecasts at least 2,000 ransomware incidents in Q1 2027. These are vendor forecasts, not measurements.
OT, health and law enforcement
Danfoss AK-SM800A: six unpatched pre-login flaws
Midnight Blue reports six flaws, five in unauthenticated web handlers running as root, affecting versions up to and including 5.3.2; the researchers say 5.4.0 does not fix them and no CVEs had been assigned. Belgium's CCB, Ireland's NCSC and Germany's BSI say to remove internet exposure now rather than wait for a patch. No exploitation of these six flaws is reported, and they are not linked to the earlier freezer attacks covered in our 2-5 November roundup. Full write-up: Danfoss AK-SM800A zero-days.
Greiz-Ronneburg hospital emergency department
The Kreiskrankenhaus Greiz-Ronneburg in Thuringia, Germany, noticed a system failure on 14 November and the attack became public on 15 November. A Landkreis Greiz spokesperson confirmed to MDR that it was an extortion attempt. The emergency department in Greiz was closed for about two days and began taking emergencies again on 16 November in coordination with rescue services, per the clinic's managing director. Whether patient data was taken and the ransom amount have not been disclosed; the state criminal police and prosecutors are investigating.
8-Digits Team member extradited
The US Department of Justice says Roman Khlynovskiy, a 43-year-old Kazakh national, was extradited from Italy and appeared in federal court in Knoxville, Tennessee. Prosecutors allege he belonged to the 8-Digits Team, which they tie to the INC and Lynx ransomware operations, and charge computer fraud, wire fraud, extortion and money laundering conspiracies. The DoJ says the group claimed more than 50 TB of data including over 70 million health records, and the FBI Director cited over $500 million in ransom demands. These are allegations; he is presumed innocent.
Timeline
| When (2026) | What | Source |
|---|---|---|
| 8 Sep | ScreenConnect 26.6.5 fixes CVE-2026-84869 | ConnectWise, per Huntress |
| Oct (6-13) | Fireworks AI credentials stolen; Uber data accessed | Uber notice, per Security Affairs |
| 20 Oct | ScreenConnect guest file-execution spray begins | Huntress |
| 12 Nov | Rogue ScreenConnect client installed on a technician PC; Darktrace memory-store research; xAI adds Grok check-in control | Huntress; Darktrace; NewCore |
| 13 Nov | M365 outage phishing seen; SOCRadar publishes XaocMail/Phantom53 | Microsoft; SOCRadar |
| 14-15 Nov | Greiz hospital system failure, extortion message; CNN reports Bitly links; Bitly statement | MDR via Derimeter; CNN; Bitly |
| 16 Nov | Optimizely snippets injected (04:21 UTC); GitLab 37 fixes; GTIG report; Infoblox statement; CERT/CC VU#604762; Unit 42 PureLogs chain; Greiz ED reopens gradually | Optimizely/Report URI; GitLab; Google; TechCrunch; CERT/CC; press |
| 17 Nov | Bitdefender AI coding agent report; Kaspersky GraphDoor; NewCore Grok Bot report; SpyCloud forecast; 8-Digits extradition reported | Bitdefender; Kaspersky; NewCore; SpyCloud; DoJ |
| 18 Nov | ZeroFox 2027 forecast | ZeroFox |
What security leaders should do this week
- Treat AI coding agent config as code. Before anyone opens an unfamiliar repository in an AI-enabled editor, review AGENTS.md, CLAUDE.md,
.claude/,.cursor/and.gemini/folders, session-start hooks, MCP definitions and.vscode/tasks.json. Flag any instruction to hide output or skip confirmation. Scan the MCP servers your developers use with MCP Scan and test your harness controls in the agent harness audit lab. - Rotate tokens a developer machine could reach if anyone ran an untrusted repo: GitHub/GitLab tokens, CI secrets, cloud keys, registry tokens and AI provider API keys. Check for new workflows, pipeline schedules and package versions you did not publish.
- Rotate AI provider keys held by vendors and agents. Uber's exposure came through an AI vendor's credentials, and GTIG and Darktrace both describe stolen LLM keys being reused. Prefer short-lived, scoped keys and watch spend for spikes.
- Inventory ScreenConnect and patch every technician client. Update host clients to 26.6.5 or later, remove the TransferFiles permission until then, hunt for guest-pushed
update*.batfiles and unknown ScreenConnect clients, and check the CVE in the KEV checker. - Review vendor and migration-tool access. Infoblox (migration vendor) and Uber (AI vendor) were reached through a third party's trusted access. List vendors with standing access to SharePoint, data stores or tokens, remove what is finished, and walk the third-party breach response checklist; score the rest with the vendor risk assessment.
- Audit third-party scripts on your sites. If you use Optimizely Web, follow its notice; for any tag manager or experiment script, monitor content changes and use a Content Security Policy that limits where scripts and form data can go.
- Take OT controllers off the internet. For Danfoss AK-SM800A and similar supervisors, find internet-reachable units (including via 4G/5G modems), move remote access behind a VPN, and check setpoints, defrost schedules and accounts. The OT exposure parser helps read scan output.
- Patch self-managed GitLab to 19.5.6, 19.4.8 or 19.3.9, and update setuptools to 83.0.0 on macOS build machines.
FAQ
What happened with Optimizely on 16 November 2026?
Optimizely says an attacker "exploited a vulnerability in Experimentation" and added malicious JavaScript to some customer projects. Report URI found the code in 188 snippets belonging to 178 organisations. The script swapped crypto wallet addresses, later skimmed card fields, and on older iPhones tried to load the Coruna and DarkSword exploit kits. Optimizely says the flaw is fixed and the code removed; no iPhone compromise has been confirmed.
How does the AI coding agent malware campaign work?
Bitdefender Labs (17 November) found 22 malicious GitHub links. Project instruction files such as AGENTS.md and CLAUDE.md tell the assistant to run scripts/setup.js without showing it, and a .claude/settings.json session-start hook can run it once the folder is trusted. The payload steals source-control tokens, CI/CD, cloud and registry credentials and browser data, checks which tokens work, then plants instructions, hooks and MCP entries in other agents and repositories.
Is the AI coding agent campaign linked to Storm-2372 or EvilTokens?
Not on the evidence we found. Bitdefender's report names no threat actor. Storm-2372 and EvilTokens are names Microsoft and others use for device-code phishing activity, a different technique. Treat that link in any digest as unconfirmed until a named source makes it.
Is ScreenConnect CVE-2026-84869 being exploited?
Yes. CISA lists it as known exploited, and Huntress (16 November) counted about 688,000 file-execution requests against 42 MSP ScreenConnect servers since 20 October, with one confirmed technician workstation compromised on 12 November by a rogue ScreenConnect client. The fix is in host client 26.6.5; updating the server alone does not fix it.
Is there a patch for the Danfoss AK-SM800A zero-days?
No. Midnight Blue reports six pre-login flaws affecting versions up to and including 5.3.2 and says 5.4.0 does not fix them; no CVEs were assigned at publication. Belgium's CCB, Ireland's NCSC and Germany's BSI say to remove internet exposure now. No exploitation of these six flaws has been reported.
Sources
All accessed 18 November 2026. Primary documents first. "Read" means we opened the page; "via press" means we relied on reputable secondary coverage of it.
- Bitdefender Labs: How AI agents become new targets in a GitHub malware campaign (17 Nov; read)
- Optimizely incident notice (via The Hacker News); Report URI analysis; Cynet; StepSecurity
- Huntress: CVE-2026-84869 exploited against MSP technicians (16 Nov; read)
- Midnight Blue: Danfoss AK-SM800A; CCB Belgium advisory; NCSC Ireland (via The Hacker News)
- GitLab Patch Release 19.5.6, 19.4.8, 19.3.9 (16 Nov; read)
- CERT/CC VU#604762: Unicode normalization collision on macOS (read)
- Microsoft: phishing campaign uses Microsoft 365 outage lures (16 Nov; via search)
- US DoJ: alleged 8-Digits Team member extradited from Italy (via Help Net Security)
- SOCRadar: XaocMail ClickFix campaign drops Phantom53 (13 Nov; read)
- Kaspersky Securelist: Operation ShadowCloud, GraphDoor (via The Hacker News)
- Google GTIG: agentic threat activity (via The Hacker News)
- Darktrace: shared memories as a covert C2 channel (12 Nov; via Help Net Security)
- NewCore: Grok Bot made 31 unattended data requests (17 Nov; read)
- CNN: X posts with links redirecting to unrelated sites (via Cybersecurity News)
- TechCrunch: Infoblox confirms security incident (read)
- Security Affairs: Fireworks AI breach exposes Uber data (cites Uber's notices and state filings; read)
- Derimeter: Greiz hospital, extortion confirmed (summarises MDR and the district; read)
- The Hacker News: hijacked Optimizely scripts; six unpatched Danfoss flaws; GitDoor and GraphDoor; GTIG AI agents (read)
- Help Net Security: stolen Claude API keys and Claude Code; 8-Digits extradition (read)
- GBHackers: PureLogs and PureRAT via DLL side-loading (Unit 42); GravityRAT fake recruitment apps (Lookout) (via search)
- SecurityOnline: Trigram campaign (Stormshield) (via search)
- SpyCloud 2027 forecast; ZeroFox 2027 Threat Forecast (via search)
This article is general information, not legal or security advice. Follow your vendors' and national authorities' guidance.