Ivanti, Palo Alto Networks, and Cisco all disclosed actively exploited flaws in network security products this week, including an Ivanti Connect Secure chain that gave attackers root access on a fully patched appliance. GhostAction hijacked about 1,900 GitHub accounts and targeted more than 30,000 repositories for secrets, and the Justice Department charged 14 people in a laptop-farm scheme that put North Korean IT workers inside more than 70 US companies.
The common thread this week is misplaced trust: a firewall at the edge, a developer account with write access, and a verified new hire each looked safe on paper. We also cover cyberattacks on grocers that pushed supermarket fridges to 50°C, and a lab test in which AI agents reached industrial equipment in about 16 minutes. Here's what you need to know.
Explore this content with AI:
ChatGPT | Perplexity | Claude | Google AI Mode
Ivanti, Palo Alto Networks, and Cisco each have an exploited flaw in a network security product
Volexity reported on October 13 that an attacker it tracks as UTA0576 chained three flaws to get root on an Ivanti Connect Secure appliance that was fully patched at the time (version 22.8R2.4). The Hacker News reports that exploitation began in early October, and Previdian's honeypots had logged 105 attempts from 50 IP addresses by October 14. In the intrusion Volexity analyzed, the attacker stole stored credentials, searched Active Directory, and sprayed passwords, and Ivanti's Integrity Checker Tool can miss the chain because the exploit changes code only in memory.
Ivanti fixed the authentication bypass and remote code execution flaw, CVE-2026-104935, in releases including 22.8R2.5 and 25.1.3.1, and CISA added it to its Known Exploited Vulnerabilities (KEV) catalog on October 13. The privilege escalation flaw is still unpatched; Ivanti says its fix will arrive "in a future release in the coming weeks."
Two more edge-device flaws landed on Wednesday. Palo Alto Networks' CVE-2026-0322 (CVSS 8.6) lets an authenticated attacker with network access to the PAN-OS management interface run code as root on PA-Series and VM-Series firewalls and Panorama, and SecurityWeek reports the company has seen "very limited exploitation." Cisco's CVE-2026-76519 (CVSS 9.8), per The Hacker News, lets an unauthenticated sender use an email attachment to write files to a Secure Email Gateway and potentially run commands as root, and there's no workaround. CISA added both to the KEV catalog on October 14, and Cisco advises reviewing logs stored off the appliance, because an attacker with root access can hide evidence on the device.
Why it matters: the devices built to protect the network are now the entry point. Upgrade to fixed releases now, and treat a clean result from a vendor integrity check as inconclusive; hunt for compromise with network and memory forensics instead. Restrict management interfaces to a trusted internal address, rotate directory credentials these devices could reach, and assume any appliance exposed before its fix shipped may already be compromised.
GhostAction hijacks about 1,900 GitHub accounts to steal secrets, and an npm flaw left 1,500+ packages open to takeover
GitGuardian reports a new wave of GhostAction, a campaign it first documented in September 2025. The Hacker News reports that attackers used stolen tokens and infostealer credentials to push a malicious GitHub Actions workflow, often named "security-audit.yml," that steals CI/CD secrets, cloud credentials, registry tokens, and AI provider API keys. Starting October 8, the campaign targeted more than 30,000 repositories across about 3,000 users and organizations. GitHub began blocking the workflow on October 9, and no threat actor has been named. Some victims were hit in both the 2025 and 2026 waves, which suggests attackers are counting on secrets that were never rotated.
Separately, Bloom Security found that npm's Trusted Publishing feature identified a linked GitHub repository only by its name, so an attacker could claim a renamed or deleted repository and publish to the original package. Help Net Security reports that more than 1,500 packages were exposed, with no known exploitation, and GitHub shipped fixes on October 2.
Why it matters: one stolen token with write access is enough to turn your CI into a secrets collector. Rotate any secret ever exposed to a repository you maintain, require two-factor authentication and short-lived, fine-grained tokens, and check recent workflow files for unfamiliar names. npm maintainers should also run npm trust list and confirm each linked repository is still theirs.
Justice Department charges 14 in a North Korean laptop-farm scheme as the FBI searches 40 Texas sites
On October 13, the Justice Department charged 14 people, including two US citizens who were arrested, ten people in China, and two unnamed defendants. CyberScoop reports that they allegedly used stolen or fabricated identities so North Korean engineers could hold remote jobs at more than 70 US companies, collecting about $3 million in pay, while facilitators hosted company laptops in US homes. The FBI also searched about 40 locations in Texas and seized 187 laptops and KVM devices.
An FBI public service announcement the same day warned that North Korean IT workers increasingly pay Americans to sit interviews, verify identities, and keep work devices connected. CrowdStrike, which tracks the activity as Famous Chollima, expects the takedown to have limited impact on the group because laptop farms are vital to its revenue.
Why it matters: this is an identity problem, not a malware problem. Verify identity throughout employment, not just at hire; the FBI recommends weighing several weak signals together, such as remote-management tools on company laptops and rotating bank accounts and addresses. Treat unexpected RMM software on a laptop as an incident.
Cyber attacks hit grocers: hijacked refrigeration at about 700 French supermarkets and an Ahold Delhaize outage
Over the weekend of October 11 and 12, attackers hijacked refrigeration controllers at French supermarkets, pushing some units to 50°C (122°F) and holding others in defrost mode for hours, which delayed alarms. SecurityWeek reports that about 700 stores were affected, at chains including Carrefour, E.Leclerc, and Intermarché, and the Paris prosecutor's office says the attack also caused fires. France's CERT-FR has flagged Danfoss, Copeland, and Dixell controllers, many of them end-of-life models. How the attackers got in is still unclear, and no group has claimed responsibility.
Separately, Ahold Delhaize took some systems offline early this week after identifying a "security issue." The Record reports that online orders and deliveries were disrupted in the Netherlands and Belgium, and the company says websites for US brands including Stop & Shop, Food Lion, and Giant also went down while stores stayed open. Ahold Delhaize has found no evidence of data theft, and both the company and the Dutch NCSC say there's no evidence linking it to the French attacks.
Why it matters: a hijacked refrigeration controller can destroy inventory, and even start fires, without touching a single laptop. Keep these controllers off the open internet, put any remote access behind a VPN or another protected network as CERT-FR advises, replace end-of-life units that can't be patched, and segment them from corporate networks.
In a lab test, AI agents moved from a network foothold to industrial equipment in about 16 minutes
Booz Allen published research last week on how frontier AI models perform against operational technology (OT). SecurityWeek reports that in Booz Allen's isolated lab, models with no prior knowledge of the systems achieved every test objective across eight scenarios. In one, a model went from a perimeter compromise to actions inside an industrial control network in just over 16 minutes; in another, it located a robotic arm, mapped its protection zones and motion limits, and moved it.
Existing defenses slowed some actions, but no single safeguard consistently stopped the models. This is a vendor lab test rather than an observed attack, so treat it as a signal about speed rather than a prediction of real-world incidents.
Why it matters: OT has long been protected by how much expertise it takes to move through it, and this test suggests AI agents can shorten that work considerably. Start with segmentation and remote-access controls, because those slow every attacker, human or machine.
Final thoughts
This week's pattern is misplaced trust. Edge devices meant to keep attackers out were used to get in, a developer account was trusted to publish only its owner's code, and a verified hire was trusted to be who they said they were. Even a refrigeration controller was trusted to sit quietly on the network.
If there's one action item, it's to verify what you've been trusting: know which internet-facing systems you run and how fast you can patch them, rotate secrets that may have been exposed, check identity beyond the hiring process, and segment anything that controls physical operations.
Check back next Thursday for the next edition of This Week in Cybersecurity.
.png)