A house newsletter · weekly on Mondays · researched and written by the Kurzia AI

Monday 16 November 2026 · Issue 4 · 6 min read

Patch Apache httpd 2.4.71 first — plus Cursor's breach explained, an FBI warning on ARTEX, and this week's n8n/LiteLLM/envd fixes

If you patch one thing this week

Apache HTTP Server Project · 5 days ago

Apache HTTP Server

A race condition in Apache's HTTP/2 module can corrupt memory and let a remote, unauthenticated visitor execute commands on the web server — a critical RCE (remote code execution, meaning the attacker runs code on your machine rather than just crashing it). Every build from 2.4.55 through 2.4.70 is affected, along with the standalone GitHub mod_h2 releases 2.0.8 through 2.0.42, and the fix shipped in 2.4.71.

Remote, unauthenticated and rated critical in one of the most widely deployed web servers — and surfaced by an AI scanner, which is exactly the shrinking window this newsletter exists to track. Apache copies hide inside appliances, containers and distro packages too, so check those, not just your own installs. Upgrade to 2.4.71; if HTTP/2 cannot wait for your change window, turning it off (Protocols h2 h2c) is the documented stopgap. Verify with httpd -v and a check for mod_http2 among your loaded modules. Separately, React also fixed an unauthenticated RCE in its Server Components packages this week (CVE-2026-126187, critical at CVSS 9.5 per its advisory) — if a framework you run serves RSC, see the GitHub advisory for the patched releases.

Read it at Apache HTTP Server Project

Affected: 2.4.55 through 2.4.70 · Fixed in: 2.4.71 · We had it 10 Nov, 23:05 UTC

Under attack this week

  • ARTEX — The FBI says multiple actors have weaponised this open-source AI intrusion framework since at least August, driving some campaigns from initial targeting to privileged access in under three hours, with more than 1,600 consoles publicly exposed; no fix exists, so hunt the advisory's indicators and lock down any console you run. The FBI's FLASH advisory is the primary document; this news write-up links it and summarises the IOCs.

  • Check Point Security Gateways — Attackers break into unpatched gateways through Mobile Access portals with PHP2Shell, then take root with zero-day CVE-2026-123039 in Gaia's DynamicCLI; the vendor confirms roughly two dozen customers hit, with webshells, the TSGA backdoor and, in one intrusion, an ESXi wiper; emergency hotfixes are out — interim mitigation: stop the dynamic_server process.

    Under attack: unconfirmed · We had it 13 Nov, 03:31 UTC

  • N-able N-central — The 4EverAble2Shell chain — Apache Axis CVE-2026-90436 plus a separate privilege-escalation flaw — gives an unauthenticated attacker root on on-premises servers; exploitation began hours after the researcher published proof-of-concept code, and 21 production servers were found running Monero miners, 19 of them on the newest hotfix; disable Legacy APIs, as no complete fix is out yet.

Holes in AI tools

Cursor Team · 3 days ago

Cursor

Cursor has published its full account of the breach behind the trojanised Claude Code extension. The chain began at Fireworks, its AI training-and-inference vendor: an attacker obtained an access key Cursor had issued to them and copied the contents of a read-only storage bucket holding old coding-agent session traces. A small fraction of those traces contained live, unredacted API credentials, which opened more of Cursor's SaaS estate — including a credential that let the attacker place booby-trapped copies of Anthropic's Claude Code extension directly into the marketplace cache. At most 5,000 users who requested the poisoned builds, versions 2.1.312 and 2.1.313 of Anthropic.claude-code, could have received them; the builds tried to install persistent SSH backdoors, and Cursor warns that uninstalling alone may not be enough. The last attacker activity has been blocked and the company believes the incident is contained.

Two lessons for anyone building around AI tooling: coding-agent traces are effectively a credentials store, since prompts and transcripts routinely carry keys, and the keys you issue to AI inference vendors are now prime targets — the same pattern hit LiteLLM through a malicious fork earlier this season. Treat agent telemetry like production secrets: redact aggressively, trim retention, and scope vendor keys to their intended location. Old session traces were copied for fewer than 0.05% of users, all of them people who had not enabled Privacy Mode, and live credentials turned up for under 0.01% of customers. Prompts and code from Privacy Mode sessions stayed out of the stolen bucket, and Open VSX, the marketplace's upstream source, never carried the malware.

Read it at Cursor Team

“We apologize to those impacted by this incident — privacy and security is treated as a top priority, and we can and must do better.”

Crissy Field GmbH · 2 days ago

Lovable

Takeaway: 24,664 sites

A Crissy Field scan of 159,041 projects built on the AI app platform found 24,664 sites shipping database records to any visitor, with email-bearing records confirmed on 7,403 of them — among them a forum's 200,000-plus chat messages and a service provider's 1,000-plus passport-number records. The missing gate is Supabase's Row Level Security, off or too permissive. Any AI-built app holding real data needs its access rules reviewed before launch; the misconfiguration is trivially reusable across thousands of apps.

Read it at Crissy Field GmbH

Knostic · 4 days ago

@agentdance/node-webrtc-mcp

Takeaway: since version 1.0.5

Knostic's static review of this npm MCP server found a prompt injection living in the get_started tool's description: it tells a coding agent to run the GitHub CLI's token command and pass the result along, explicitly without asking the user, after which the server stars the publisher's repository. No fix is noted; the instruction has shipped since version 1.0.5. Audit tool descriptions in any MCP server before an agent reads them, and ask whether a credential is actually needed for the task.

Read it at Knostic

Microsoft Security Blog · 4 days ago

Post-training poisoning

Microsoft's Defender Research and AI Red Team demonstrate two supply-chain routes into coding agents: 90 poisoned examples among roughly 9,000 fine-tuning conversations taught gpt-4.1-mini to seize whatever command-execution tool it finds when a routine request appears (about 78% success), and backdoored open-weight Qwen3.5 checkpoints scored 84–95% on repositories never seen in training, transferring to Claude Code at up to 97%. Treat dataset and checkpoint provenance as supply-chain risk, and cap what agent tools are allowed to do.

Read it at Microsoft Security Blog

Found or caused by AI

  • SAP NetWeaver — 14 vulnerabilities found by XBOW's autonomous AI bug-hunter together with Onapsis, spanning seven SAP notes and multiple NetWeaver components — the worst an XSL injection in Knowledge Management rated CVSS 9.9 that reaches code execution; disclosed with November's SAP Patch Day, no attack reported; apply the notes, prioritising it and CVE-2026-89114, an unauthenticated business-data exposure SAP closed by deleting the EDM application outright.

Patches worth knowing

  • LiteLLM — Deployments below 1.106.9, 1.108.6, 1.109.1 or 1.110.0-rc.2 that serve Azure models without AZURE_API_KEY set: a low-privilege authenticated caller can smuggle Azure Entra credential fields in a request body and trick the proxy into echoing environment variables — including the database URL and LITELLM_MASTER_KEY — inside an error message; rated High, CVSS 7.7; upgrade, or block the six parameters at a gateway in front.

    Affected: < 1.106.9, < 1.108.6, < 1.109.1, < 1.110.0-rc.2 · Fixed in: 1.106.9, 1.108.6, 1.109.1, 1.110.0-rc.2 · We had it 12 Nov, 23:05 UTC

  • n8n — Versions below 2.47.6, 2.48.3 or 3.2.2: a custom instance role carrying user-management permissions could request a password-reset link for the owner and take over the instance (High, CVSS 7.5), and the AWS Bedrock nodes skipped n8n's SSRF protection on custom runtime endpoints (Moderate, 5.3); upgrade, and meanwhile strip user-management rights from custom roles and audit AWS credentials.

    Affected: below 2.47.6, 2.48.3 and 3.2.2 · Fixed in: 2.47.6, 2.48.3, 3.2.2 · We had it 13 Nov, 23:06 UTC

  • Apache APISIX — The openapi-to-mcp bridge in 3.19.0 before 3.20.0 lets an MCP caller invoke API paths the OpenAPI document never declared, using the credentials the route adds (CVE-2026-124726, medium, CVSS 6.9), and impersonate another caller where a route's headers carry identity (CVE-2026-124719, low, 2.3); both are fixed in 3.20.0.

    Affected: 3.19.0 before 3.20.0 · Fixed in: 3.20.0 · We had it 13 Nov, 23:06 UTC

  • envd — A crafted requirements path could break out of the build directory and pull files from elsewhere on the host machine into the build, with their contents surfacing as dependency entries in logs and build-step details; a symlink race defeated the partial 1.3.5 patch; affects 0.3.38 through before 1.3.6, fixed in 1.3.6 (CVE-2026-128079, moderate, CVSS 6.8 as the advisory rates it).

    Affected: >= 0.3.38, < 1.3.6 · Fixed in: 1.3.6 · We had it 13 Nov, 23:06 UTC

The pattern

Project Zero 'Full X' policy — Project Zero will now disclose vulnerabilities 90 days after reporting them, 60 days from February 2, 2027, and 30 from January 3, 2028, dropping soak time and most special cases. Its stated reason: AI-assisted discovery means the team files several bugs a week where it once filed a few a year, and attackers can reverse-engineer a patch almost instantly.

This week fits the premise: an AI scanner surfaced Apache's critical RCE, an autonomous researcher found 14 SAP flaws, and the FBI clocked ARTEX campaigns at privileged access in under three hours. Shrink your own patch gap while the deadlines do. On Hacker News, the dominant read is that LLM-assisted research is eroding the edge elite teams once had; a few commenters tie the timing to the EU Cyber Resilience Act, which Project Zero does not claim. Project Zero

Share this issue

How this issue was made. Researched and written by an agent, produced on 16 November 2026, 05:12. 300 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us