The most instructive story of the week is not the wiper. It is the socket.
Check Point published an advisory last week for CVE-2026-123039, a local privilege escalation flaw in the DynamicCLI daemon on Gaia OS. The daemon runs as root. It listens on a world-writable socket at `/tmp/dynamic_socket`. It does not authenticate the processes that connect to it. Check Point says it observed active exploitation against "a couple of dozen customers" since late October.
That is not a subtle bug in a hard-to-reach code path. That is a design decision, made years ago, that convenience mattered more than isolation. It sat there quietly until someone found a way in.
The Chain That Matters
The way in was PHP2Shell. Two unauthenticated remote code execution flaws in core PHP, CVE-2026-109618 and CVE-2026-109619, disclosed on 21 October and patched the same day. They affect any PHP installation serving web requests: PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP, LiteSpeed. The PHP team noted that mod_php carries the highest risk, because an unauthenticated request can trigger the attack before any application code runs.
Then the timeline gets uncomfortable. Amazon Threat Intelligence found exploitation attempts starting four days before the patches shipped. CISA added both to the KEV catalog on 26 October.
On Check Point gateways: PHP2Shell against the Mobile Access portal for initial access, CVE-2026-123039 for root, webshells and the TSGA backdoor for persistence, then a pivot to VMware ESXi hosts found with the open-source gogo scanner, then a wiper called `destroy_v6` that zero-fills target directories before deletion to make recovery harder.
The attacks began on 30 October against Israeli organisations, and a second wave on 2 November hit companies in several countries and sectors. Check Point tentatively linked the tooling to PhiliKit, previously seen against Ivanti appliances and which ESET associates with the UNC5221 cluster, while saying attribution could not be made with confidence. On the ESXi hosts, the attacker left a note called `DREAM_ANNEXATION.txt` carrying political messages and explicitly refusing ransom negotiation. Whatever this was, it was not extortion.
Siemens is still working through which of its products ship a PHP runtime, and published ICS advisories last week covering building automation, energy management and industrial edge products against the same two CVEs. This is what a language-level flaw does in a world of embedded software: it stops being one advisory and becomes dozens, each with its own patch schedule, each owned by a different team.
Two Vendors, Two Ways of Failing
N-able had a worse week. A researcher, Sean Bales, published more than 30 alleged zero-day vulnerabilities in N-central, the remote monitoring and management platform MSPs use to administer customer estates. N-able took its hosted environments offline and told on-premises customers to shut down their servers. It released emergency hotfixes on 10 November. Bales published working proof-of-concept exploits for two more flaws on 12 November, tested against that hotfix, and researchers reported at least 21 compromised systems within hours, including 19 running the latest update. The exploits were being used to deploy cryptocurrency miners.
There is a version of this story where the vendor is the victim. That is not the version I would tell a board. N-able shipped a fix that did not close the class of flaw, then had its platform taken down by a researcher who had lost patience with coordinated disclosure. Bales said N-able had not patched two unrelated RCE vulnerabilities he disclosed in July, and his complaint is specific and checkable.
ConnectWise is the cleaner case, and therefore the more alarming one. The ScreenConnect attacks compromised at least 288 endpoints across 48 organisations, with every affected device hit inside 90 minutes and roughly a third inside the first 10. Huntress traced initial access to RMM+, a third-party extension, versions 1.0.31 and 1.0.32, which used a shared access key rather than ScreenConnect's own authentication. If an administrator had never configured that key, it stayed blank. An attacker could supply an empty key, pass the check, and issue remote commands without logging in or completing MFA.
A blank password was accepted as valid authentication, and it bypassed multifactor entirely, inside the tool MSP technicians use to hold the keys to their clients' estates. ConnectWise disabled five potentially related third-party extensions across roughly 400 partner environments.
The follow-on toolkit, SCKit, is a collection of 16 scripts built to spread from an infected customer device to a technician's workstation and onward into other managed networks. It attempts to exploit CVE-2026-84869, a previously patched ScreenConnect flaw. Huntress and Blackpoint both say they have not seen that exploit succeed, and Blackpoint's testing against a patched client failed. That shows the worm was written for a world where patching lags.
Machine Access Is Now the Attack
On 10 November at 12:45 UTC, an actor used credentials from 4,837 compromised GitHub accounts to probe 61,764 repositories for write access. They opened pull requests titled "docs: dependency audit notes". Aikido, which documented the campaign under the name wfx-loot, reports that 92% of those pull requests remain open, which tells you how closely anyone was watching. Where write access existed, a 1,500-line credential-stealing script went in through a modified GitHub Actions workflow, a new workflow on a short-lived branch, or an npm install hook.
The script harvested GitHub Actions and OIDC tokens, AWS, Azure and Google Cloud credentials including instance metadata on self-hosted runners, and secrets from Git, SSH, Docker, Kubernetes, npm, Vault and Terraform. It also collected credentials for OpenAI, Anthropic, Gemini, Hugging Face, Cursor and GitHub Copilot. It exchanged OIDC tokens for temporary cloud sessions, chained into IAM roles to mint long-lived credentials, and impersonated GCP service accounts. Then it reverted its changes and deleted the logs inside a minute.
Some of the compromised accounts belonged to Fortune 100 organisations, the US General Services Administration, the French government, Nextcloud and WebdriverIO. Roughly 50 had publishing rights on npm, PyPI, Go, crates or Packagist. Automated publishing pushed the payload into Packagist's `digitalelvis/neuronai-studio` 3.14.3 and two versions of the Go module `github.com/onuragtas/docker-env`.
The pattern repeats. TSMC disclosed to the SEC that a cloud service provider had been breached, exposing data on current and former employees. No TSMC systems were compromised; the data was verified to match records held in the provider's systems, and the provider was not named. A forum post on 8 November offered 6.9 GB of employee data, including banking details, for 3 BTC, roughly 200,000 dollars. TSMC supplies Nvidia, Apple, Broadcom and Qualcomm and controls more than 70% of the semiconductor market. Its security posture held. Its supplier's did not, and the employees carry the consequence.
What the Regulator Sees
The ICO recorded 1,371 cyber incidents between July and September, the highest quarterly figure on record and above the previous 2023 high. Total data security incidents reached 4,604 for the quarter. Ransomware reports, by contrast, are falling: 1,253 in 2023, 752 in 2024, 620 in 2025, and just 321 in the first nine months of 2026.
That divergence is the story, not a curiosity. Sophos counted more than 1,200 new victims named on leak sites worldwide in October, its highest monthly total and nearly two thirds more than a year earlier, driven mainly by Qilin, The Gentlemen and n0n. Around 350 British organisations have been named. Encryption is declining while extortion rises, because stealing the data is cheaper and more reliable than holding a file system hostage. If your incident response plan is built around decrypting backups, it answers last decade's question.
Where AI Actually Bites
The AI agent story I would take to a board is not an attack at all. It is a Cloud Security Alliance survey commissioned by Token, covering 418 IT and security professionals, in which 68% claimed strong visibility into their AI agents while 82% had found agents in the past year they did not know existed.
Consider how those agents get built. Someone in accounts connects a no-code builder to email through an OAuth consent screen, generates a personal ERP API key by following the model's instructions, and attaches Slack through the builder's connector. Nobody in security is involved. Nobody needs to be. The agent reads an inbox and posts a summary, and it holds credentials with more reach than most of your service accounts.
That is the gap the GitHub campaign exploited: credentials held by something never registered as an identity. The answer is short-lived, purpose-built agent identities, the fourth of the four access paths practitioners describe and by common account the rarest.
The Week in Numbers Worth Knowing
Two enforcement results change the calculus. Operation Endgame took down 235 servers and 257 domains across Gootloader, CastleRAT and StealC. Gootloader has operated since 2018, infected 25,000 devices, and according to the DOJ fed at least 44 US ransomware victims for losses of 28.8 million dollars. German lawmakers also passed the largest overhaul of the country's intelligence laws since the war, 308 votes to 190. The BND gains powers to act against threats abroad rather than merely report on them, including sabotage of adversary supply chains and takedown of hostile servers. For anyone operating in Germany, that is a new actor in the environment, subject to a different legal framework than you are used to.
What To Do This Week
Inventory every internet-facing Check Point gateway and install the November Jumbo Hotfix. Prioritise anything exposing the Mobile Access portal. Where you cannot patch immediately, stop the `dynamic_server` process and prevent it restarting. Then hunt: session files containing `upload_progress_pwn`, unexpected `CPconfigdaemon` processes, unexplained gaps in firewall logs, and SSH from a gateway to internal ESXi hosts. That last one is a breach, not a curiosity.
Sweep for PHP2Shell outside your web servers too. The patch is PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35, and the harder problem is appliances, building systems and embedded tooling carrying a PHP runtime nobody knows about. Ask facilities and ICS vendors to confirm their status against the Siemens advisories, and do not accept silence.
Treat RMM as tier zero. Every ScreenConnect, N-central and comparable agent is a privileged path into everything it manages. Confirm which extensions are installed, remove anything you cannot justify, and verify no access key is blank. Where an extension authenticates with a shared static key, that key sits outside your identity provider, invisible to MFA. Find them and move them into something that expires.
Assume your CI is a credential store and govern it accordingly. The GitHub campaign succeeded because a pull request with a plausible name got workflow write access. Restrict who can approve workflow changes, require review on `.github/workflows`, and audit your OIDC trust policies this week. If a repository can exchange a GitHub token for an AWS role, that path is an identity and it needs an owner.
Get your AI agents into the asset register. Ask for every agent touching email, ERP, finance or Slack, including those built by business units on personal API keys. You will not get a complete list, so plan for that.
Rewrite the playbook for extortion without encryption. Falling ransomware numbers and rising leak-site counts mean the case you are most likely to face involves stolen data and a deadline, not locked files. Check that your legal, communications and negotiation plans work when nothing is encrypted and there is nothing to restore.
The wiper on those ESXi hosts got the headlines. The world-writable socket that let it happen had been there for years, waiting for someone to notice. Most of what was exploited this week was not novel. It was old, unowned and unexamined, sitting somewhere nobody had thought to look.

