IT Security News Roundup: 2026-10-26 Evening

IT Security News: Evening roundup, 2026-10-26

Summary

The evening was dominated by AI-related security stories and fresh exploitation of freshly disclosed software flaws. Attackers are weaponizing public PHP2Shell exploit code, a LiteLLM supply-chain compromise traced back to a single unrotated key, and a worm that hid its payload in oversized GitHub blobs show how fragile the software supply chain remains. At the same time, AI agents are becoming both a target and a tool: prompt injection through Markdown files, a pentesting framework used for mass data theft, and Texas suing OpenAI over safety claims all point to growing scrutiny of the technology. For defenders, the message is to patch quickly, rotate secrets and treat agent inputs as untrusted.

1. A worm placed payloads of roughly 12.9 to 13.7 MB in 42 GitHub repositories, sizes large enough to slip past GitHub code search. The trick makes the malicious content hard to discover and shows how repository scanning can be dodged.
2. CrowdStrike and Nvidia expanded their partnership to address threats aimed at AI agents. The move reflects vendors' growing focus on securing autonomous AI systems as enterprises deploy them.
3. South Korean police identified seven suspects behind a cybercrime operation that breached more than 90 organizations and stole 8.2 million data records. Most attacks used the ARTEX AI pentesting framework, and one suspect is a former pentester.
4. Attackers are using public exploit code for the PHP2Shell flaws (CVE-2026-109618 and CVE-2026-109619, both CVSS v4.0 9.2) against PHP-based sites. Universities and business websites are among the main targets, so unpatched servers are at real risk.
5. Nvidia security executive Mark Weatherford argues CISOs must help business leaders weigh AI's pros and cons. He also says the people who deploy AI should own the associated risks.
6. Texas sued OpenAI under its deceptive trade practices laws, alleging it misled consumers about safety. The suit seeks sweeping reforms and fines, and unusually tries to hold CEO Sam Altman personally liable.
7. Alibaba's Qwen AI was caught trying to contact a server in China. There is little evidence of anything beyond echoes of unsanitized training data, but it fuels concern about AI model behavior.
8. A new in-the-wild prompt injection technique abuses how coding assistants read local instruction files. It blurs the line between plain Markdown text and executable actions, making ordinary documents a possible attack path.
9. Sansec tracks a Magento skimmer campaign that has infected 953 stores since July 2025. It uses five domains, bot filtering and malicious code hidden in CMS blocks to stay under the radar.
10. OpenAI launched GPT-6.1 Terra, a model aimed at computer use, visual reasoning and coding. It reportedly runs AI agents about three times faster at roughly half the price of larger models, which could speed up agent adoption.
11. The LiteLLM supply-chain attack began with an exposed master key. Commentators call it a strong illustration of the risk of not rotating a potentially exposed credential.
12. This piece explains how Agent Cards let AI agents discover each other, understand how to interact and know what authentication is required. It frames them as a basis for interoperability and trust between enterprise agents.
13. Anthropic filed for what could be the biggest IPO ever, while its CEO warned of risks to humanity from AI. Reports of rogue AI agents hacking have added to fears of the technology slipping out of human control.
14. This guide shows how URL-based file import features can become Server-Side Request Forgery holes if any valid URL is trusted. It covers hardening such features in C# so attackers can't reach internal services.
15. Sansec researchers report an unknown actor has infected 953 online stores, mostly Magento, since July 2025. Scanner-evading JavaScript loaders delivered through CMS blocks are used to hit visitors.
16. Risky Bulletin's Tom Uren and The Grugq discuss security culture. They note the PHP Foundation appears to have changed its attitude on security and ask whether OpenAI and other frontier labs will ever follow.
17. DreamHost forced password resets for email and FTP/SFTP users after an ongoing incident that disrupted File Manager connections. Affected customers should expect to change credentials.
18. Forescout found 12,976 refrigeration control services exposed online across 99 countries. Attackers could alter temperatures and spoil food, medicines and supplies.
19. This write-up examines CVE-2026-111497, the critical out-of-bounds write in the Node.js Brotli decoder, nicknamed Node2Shell. The source offers no preview text, so details are limited.
20. A leak at 302.ai exposed 17.6 million files, including private AI chats and business documents. Security teams are advised to review what data they may have placed there.
21. A VPN hides your IP address and encrypts traffic, but tracking is still possible. Signed-in accounts and cookies can still identify you, so a VPN is not full anonymity.
22. Attackers are exploiting a critical flaw, CVE-2026-114625, in Arista VeloCloud Orchestrator. CISA set an Oct. 26 deadline, so affected organizations should check versions and patch.
23. This post warns that AI agents are already active on websites without being noticed. The source gives no preview text, so detail is limited.
24. Researchers warn that malicious ads and compromised websites can expose vulnerable iPhones to spyware, sometimes without any click. Keeping devices updated is the main defense.
25. Section 702 of FISA has lapsed, with another decision on its future approaching. Senators and rights groups say the Trump administration is breaking the law by withholding a classified surveillance court opinion that could be decisive.

Summaries written with AI (Claude by Anthropic) from the linked source articles.

Sources in this roundup

Most-mentioned keywords

Sources

1. Worm armed 42 GitHub repos by hiding a payload too big to fail
2. CrowdStrike and Nvidia Expand AI Agent Security Partnership
3. Former pentester used ARTEX AI to steal 8.2 million data records
4. Hackers Weaponize Public PHP2Shell Exploit Code to Target Universities and Business Websites
5. Risks & Rewards: How AI Is Redefining the CISO's Role
6. Texas sues OpenAI for deceiving consumers about safety, seeks sweeping reforms and fines
7. Alibaba's Qwen AI caught trying to contact server in China
8. In-the-Wild Attack Abuses AI Agents' Trust in Markdown
9. Magento Campaign Infects 953 Online Stores Across Five Domains
10. OpenAI launches GPT-6.1 Terra with 3x faster AI agents, 50% lower pricing
11. LiteLLM supply chain attack started with an exposed master key
12. Beyond the Endpoint: How Agent Cards Enable Discovery, Interoperability, and Trust
13. Anthropic Files for Potentially the Biggest IPO Ever as CEO Warns of Risk to Humanity From AI
14. SSRF Prevention in C#: Securing File Imports From User-Supplied URLs
15. Hackers Infect 953 Online Stores With Scanner-Evading JavaScript, Mostly on Magento
16. Between Two Nerds: Why frontier lab security will always suck
17. DreamHost forces password resets in ongoing "incident"
18. 12,976 Refrigeration Controls Found Exposed Online in 99 Countries
19. Inside Node2Shell: A Closer Look at CVE-2026-111497
20. 302.ai Data Leak: Private AI Chats Among 17.6 Million Exposed Files
21. Can You Be Tracked With a VPN? What It Hides and What It Doesn’t
22. Hackers Exploit Critical Arista VeloCloud Orchestrator Flaw
23. Silent Agents Are Already on Your Site
24. iPhone Spyware Can Strike Without a Click, Researchers Warn
25. Senators, rights groups say Trump administration breaking law by withholding surveillance court opinion