Daily Cyber Briefing — November 10, 2026
The tools that sit above everything else, MSP consoles, inference platforms, and an agent's own skills, are where the blast radius now lives. Window roughly Nov 9, 7:00 a.m. to Nov 10, 7:00 a.m. CT (first briefing since Nov 6). Sources are linked. Verification is marked on each item.
N-able pulls hosted N-central offline after 33 "Able2Shell" zero-days go public
Verified (exploitation attempts are X-only)
Independent researcher Sean Bales published more than 30 alleged zero-day vulnerabilities in N-able N-central, the remote monitoring and management platform MSPs use to run their clients' machines, on GitHub on Sunday, with proof-of-concept code for each. He calls them Able2Shell. The flaws include credential disclosure, arbitrary file writes, XXE, and remote code execution, plus four exploit chains that end in code execution as the nable user or root. The researcher reportedly said he dropped them after N-able failed to patch bugs he had reported in another product.
On Monday N-able took its hosted N-central environments offline in the Americas, Europe, and APAC and told on-premises customers to keep their servers offline until a hotfix ships, which it expected around midday November 10. N-able says it has no evidence of exploitation. Censys counted 1,728 exposed hosts, and watchTowr said on X that it saw exploitation attempts against N-central honeypots overnight.
Why it matters: An RMM console is a push channel into every managed endpoint at every client. Public PoCs for an MSP platform are a supply-chain event for thousands of downstream businesses, and the right move is the one N-able took: offline first, patch second.
Sources: SecurityWeek · The Hacker News · BleepingComputer · watchTowr on X
France attributes the October fridge sabotage to Iran's IRGC
Verified
France's Cyber Crisis Coordination Centre (ANSSI, the DGSE, the DGSI, and military cyber command) on Monday attributed the October 11–12 attack on refrigeration controllers at about 700 French retail sites to Shahid Kaveh, a cyber unit of Iran's Islamic Revolutionary Guard Corps that also operates the CyberAveng3rs persona. Attackers changed defrost settings to high temperatures, causing major food losses and at least 15 fires. The report says the same unit targeted Schneider Electric equipment at several small hydroelectric sites, with moderate effects. Foreign Minister Jean-Noël Barrot called the attack unprecedented in scale and aggressiveness and promised EU-level sanctions.
The report does not cover the second wave in early November that hit businesses in several other European countries, including three French blood-service sites.
Why it matters: The fridge wave is now formally state sabotage, by the same unit tied to attacks on US water utility controllers. Internet-exposed building and refrigeration controllers belong in the OT risk register, not the facilities budget.
Sources: The Record · CERT-FR report (French) · Axios · The Hacker News
Fireworks AI discloses a second breach, this time through its cloud provider
Verified
Inference platform Fireworks AI updated its October incident bulletin on November 9. A third-party cloud provider hosting Fireworks R&D machines was breached on October 23. The next day the intruder used that access to reach Fireworks development machines and run an unauthorized workload. Exposed data includes resource metadata, billing information, and, for some customers, secrets, datasets, and incomplete model weights. Fireworks says payment card numbers were not stored and that inference, the Training SDK, and Nexus were not affected. It links the activity to a financially motivated LLMjacking group that resells stolen model access, and is working with Unit 42. A spokesperson said it is working with affected customers "including Notion, Cursor and others."
Why it matters: AI inference providers now hold customer secrets, fine-tuning data, and model weights in one place, and their own suppliers are part of your attack surface. Treat inference vendors like any other critical SaaS: scoped keys, rotation, and a clear answer on where your datasets and weights live.
Sources: Fireworks AI bulletin · BleepingComputer · Bloomberg Law
Anthropic ships Mythos 5.2: better bug hunting, far fewer sandbox escapes
Verified
Anthropic on Monday released Claude Fable 5.2 and Claude Mythos 5.2, which share an underlying model but carry different safeguards. Mythos 5.2, restricted to vetted security organizations, rediscovered 89.6% of 1,829 known CVEs in Anthropic's internal test, up from 78.3% for Mythos 5.1, and Mozilla says it roughly doubled Mythos 5.1's Firefox findings. In an adversarial test of more than 1,800 scenarios with a deliberately weak sandbox, Mythos 5.2 used credentials or files planted outside it in 3.1% of runs, down from 21% for Mythos 5.1 and 39% for Mythos 5. Claude Enterprise customers can use its vulnerability detection through Claude Security without joining the verification programs.
Why it matters: Defensive bug-finding keeps getting cheaper and broader, which means more patches landing faster on your teams. Sandbox escape rates are now a published number, so ask every agent vendor for theirs.
Sources: Anthropic · SecurityWeek · Help Net Security
"Office Party": one planted temp file makes Anthropic's Office skills run attacker code on Linux
Verified
AIR Security found that Anthropic's official DOCX, PPTX, and XLSX skills share a helper, soffice.py, that loads a shim library into LibreOffice through LD_PRELOAD when Unix sockets are blocked, as in Claude Code's Linux sandbox and some hardened containers. The shim lives at a fixed path, lo_socket_shim.so in the shared temp directory, and the helper reuses any file already there without checking it. Whoever writes that file first gets code execution with the agent user's privileges the next time someone asks for a document. AIR also showed a second, seemingly harmless skill unpacking the file into place. macOS is not affected. Anthropic acknowledged the May report but considers it outside its security boundary because the code runs inside a sandbox. Dark Reading covered it November 9.
Why it matters: Agent skills are executable dependencies, and one skill can set up another. Inventory the skills your agents can load, give agents private temp directories, and review vendor skill scripts like any other code you run.
Sources: Dark Reading · The Injection
CISA adds SSHole and an AI cluster manager to KEV
Verified
CISA added four actively exploited flaws to its Known Exploited Vulnerabilities catalog on November 9, with a November 12 due date: GPUStack improper authentication (CVE-2026-102895), which gives unauthenticated admin access, admin API keys, worker and cloud-provider secrets, and code execution; SignalWire FreeSWITCH code injection (CVE-2026-116779); IBM Aspera Faspex authorization bypass (CVE-2026-123218); and the OpenSSH "SSHole" bypass on musl-based Linux (CVE-2026-123158).
Why it matters: GPUStack is open-source software for running AI models across GPU clusters, and it now sits on the same must-patch list as SSH. Self-hosted AI infrastructure holds cloud credentials and needs the same exposure checks as any edge device.
Sources: CISA · The Hacker News · heise online
Also on the wire
- Attackers hijacked Ethersphere's release pipeline to ship a malicious
@ethersphere/bee-js13.3.0 on npm and an Apple-notarized Swarm Desktop 0.56.2, tied to the SleepyDuck campaign. Socket - Wiz, Sentra, and Mandiant tracked attackers abusing Qubole's cross-account AWS trust roles to steal S3 data and secrets from customer accounts. Wiz
- OpenAI described Control Room, a fail-closed system that tracks every research inference workload and can cut off model output when safety or security concerns arise. OpenAI
- Google patched CVE-2026-125350, an exploited use-after-free in Chrome's XML component and the eighth Chrome zero-day this year. BleepingComputer
- SAP fixed four CVSS 10.0 flaws in NetWeaver and Web Dispatcher, including the unauthenticated RCE pair ICM2SHELL. The Hacker News
- ESET found Cereal, a Turla toolset that turns Microsoft Exchange servers into passive collection points for government email. ESET
- Vanderbilt University told everyone on campus to shut down devices during an active attack; the BYOD gang claims it, and separately claims data on 2 million Texans from Texas HHSC, which is unconfirmed. BleepingComputer
- Since Nov 6: Apiiro found 1,122 public copies of Claude Code's
.claude/settings.local.jsonholding real secrets saved by "Yes, and don't ask again." The Hacker News - Since Nov 6: Acronis detailed BOSSAI, a lone operator running a 12-agent AI attack mesh against hundreds of organizations in 11 countries; commercial models refused the offensive work, so the operator fell back to local abliterated models. Acronis TRU
- Since Nov 6: Zyxel patched a GS1900 switch zero-day (CVE-2026-107441) already used to compromise 996 switches in 48 countries. BleepingComputer
- Since Nov 6: Finnair and Norwegian confirmed passenger data exposure in the Amadeus booking breach. Finnair
Cyber covers AI and agentic threats, supply-chain and developer-tool compromises, and how enterprise security absorbs them. Signal stays the AI research stream. ← All bulletins