DeepSeek and Microsoft Feature in AI Security Research

Key Takeaways

• Palo Alto Networks says Cairn, paired with Zhipu GLM and DeepSeek, was used in 25 operations against at least 19 organizations; attackers accessed or retrieved data from 10.
• Unit 42 found 1,323 internet-facing CyberStrikeAI hosts and 609 ARTEX hosts as of October 2026; exposed password-reset pages and APIs enabled some cloud intrusions in under a day.
• Microsoft’s AI-assisted Linux scan yielded 59 public bug reports from 26,502 potential findings; eight flaws had working local privilege-escalation exploits, and 38 reports had fixes in mainline at publication.
• Adobe showed creative agents that can assemble vlogs or ads in minutes, while saying users can choose AI or non-AI workflows; Firefly uses content Adobe has permission to use.
• Google and Highsnobiety’s PIFT showcased designers using Pixel and Gemini for product development, research, digital silhouette testing and marketing images.
• Medical-regulatory experts say specialized AI can speed research, but human review and final decisions remain essential.
• A survey of more than 800 workplace-learning professionals found 75% saw faster content creation, but only 11% reported better job performance and 35% consistently measured it.
• Taoyuan reports 20% fewer complaints about modified exhausts and ambulance travel times falling from four minutes to just over two when electronic systems create green-light routes.
• Articles on education and security testing stress meaningful choice and human oversight: students should be able to refuse AI, and AI refusals in IT/OT tests can leave dangerous gaps while creating false confidence.
• Erin McGee Ferrell’s “Robots Y’all” exhibition at Garner Narrative explores AI-related hopes, fears and deception and remains open through Dec. 13.

AI Agents Used Cairn in Attacks on 19 Organizations

From August to September 2026, Chinese-speaking attackers used the open-source Cairn penetration testing framework with the Chinese AI models Zhipu GLM and DeepSeek. Cairn planned and carried out 25 operations against at least 19 organizations, and attackers gained access or retrieved data from 10. They accessed credentials, exposed more than one million personal records at one organization, retrieved military and government identities, created administrator accounts, and changed a live industry standards-ballot document. Palo Alto Networks assesses with moderate confidence that the campaign was China-linked intelligence collection and tracks it as CL-STA-1238, targeting advanced semiconductor manufacturing, U.S. defense policy, and civil society.

AI Red Team Tools Let Attackers Run Intrusions at Scale

Unit 42 reports that attackers are using AI-powered red teaming tools such as ARTEX and CyberStrikeAI to carry out operations with limited human input. The team investigated intrusions involving ARTEX and tracked a campaign against customer cloud environments in sectors including retail and consulting. As of October 2026, Unit 42 found 1,323 internet-facing hosts running CyberStrikeAI and 609 running ARTEX, with 21 running both. The article says attackers used exposed password reset pages and APIs to reach cloud environments, sometimes in under a day and without exploiting a software vulnerability. It recommends that defenders continually find and secure internet-exposed assets, and test their own systems with offensive tools.

Phoenix Hill Exhibit Uses Robot Art to Explore AI Fears

Artist Erin McGee Ferrell’s “Robots Y’all” exhibition explores people’s hopes and fears about artificial intelligence. She created the work after two family members were scammed online and drew inspiration from a “Robot Repair Shop” installation at Pittsburgh International Airport and Japanese tin toy robots from the 1950s. Ferrell also built a large robot called “Saturday” with help from four strangers, using materials such as aluminum pans and chicken wire. The robot greeted visitors with a motion-sensor voice, and the exhibition at Garner Narrative is open through Dec. 13. Ferrell says AI helps with her business and marketing work, but she also wants the show to open discussion about the technology’s ability to deceive people.

Google Pixel and Highsnobiety Show AI Fashion Projects

Google Pixel and Highsnobiety’s Pixel Institute of Fashion and Technology, or PIFT, paired designers with technology for projects shown on Nov. 12. Designers used Pixel and Gemini for product development, collection research, and marketing images. OBS co-founder Matthias Schweizer used Gemini to develop a heavy cotton-canvas bag, while Chet Lo used Gemini’s Deep Research to explore shadows and 1980s anime for a capsule collection. Lou de Bètoly used Pixel to make animations, and Krobos founder Lukas Krob developed a dress from the shape of a flower, testing silhouettes digitally. Google said PIFT aims to explore what creative people do with its technology.

Writing Classes Should Teach Students They Can Refuse AI

In a Nov. 12, 2026 article, Loren Gaudet argues that teaching ethical AI use in writing classes should include students’ right to refuse generative AI. Gaudet says educators should look beyond whether students can use tools such as LLM chatbots responsibly and ask whether their use is ethical in the first place. The article says writing involves choices about how people treat readers, and classroom AI rules can shape students’ habits and values. It warns that relying on chatbots for discussion could discourage students from seeking help from peers, writing centers, teachers, or librarians. Gaudet calls on educators to give students reasons and opportunities to resist or decline AI use.

Adobe opens creative tools to more people with AI agents

At its MAX conference in Miami Beach on Nov. 12, 2026, Adobe showed creative agents that can assemble a vlog or an ad in minutes. Adobe chief Lara Balazs said the company updated its mission about a year ago to invite everyone to create, including people without design training. Adobe says its Firefly model is commercially safe because it uses only content Adobe has permission to use, and the app also offers outside models and tools from Topaz Labs. Balazs said people can choose whether to use AI, and Adobe will celebrate work made without it.

Specialized AI can speed medical device regulation, experts say

At the RAPS AI in Regulatory Conference in Barcelona on Nov. 12, 2026, experts said specialized AI can help medical device companies handle regulatory work while saving time and money. Sander Denissen of Guideways B.V. said AI agents can process large amounts of data and search, but people must set goals, check results, and verify device information. Liz Munro of Liz Munro Consulting said Guideways AI helps her research regulatory strategies for early stage medical technology companies, leaving more time to advise clients. She also warned that AI judgment has limits and people must make final decisions. Denissen said human review can take hours or days, but is needed to prevent errors and protect sensitive documents.

Clear goals and better data help turn AI training into results

A survey of more than 800 workplace learning professionals in over 25 countries found that AI often helps teams make training content faster, but that does not always improve job performance. While 75% cited faster content creation as a clear effect of AI, only 11% reported improved on the job performance, and just 35% of learning teams consistently measure that outcome. More than one in five said a lack of agreed success measures was their biggest barrier to proving AI’s impact. The report also found that 73% of teams given a leadership order to use AI still need permission to test new tools, while 53% named organizational limits such as missing measures or data access as barriers. The article says leaders should agree on intended results and how to measure and reward them, rather than focusing only on speed.

Taoyuan shows AI’s benefits and risks in government

Taoyuan, Taiwan, uses AI to identify illegally modified vehicle exhaust pipes, watch construction sites and factory chimneys for pollution, and adjust traffic lights. The city says complaints about modified exhaust pipes fell 20% over the previous year, while ambulance travel times dropped from four minutes to just over two when electronic systems create green light routes. The article warns that governments are also using AI predictions for decisions about matters such as policing, sentencing, jobs, and government services, even when the systems cannot explain their results. Predictive models can miss important context, misclassify people, and lead to harmful decisions. The article says these systems raise technical and ethical concerns and can threaten people’s rights and opportunities.

AI refusals can leave gaps in IT and OT security tests

A webinar scheduled for Wednesday, Nov. 18, from 17:00 to 18:00 CET will discuss how AI refusals can weaken security testing in IT and operational technology environments. Cracken cybersecurity scientist Dr. Oleksii Baranovskyi will explain how skipped test steps may be reported as having no findings, creating false confidence while paths into operational technology remain untested. The article says refusals often affect tests involving credential abuse, lateral movement, and industrial control system protocols. Repeating prompts can waste time and produce inconsistent results, which can undermine audits under NIS2, DORA/TLPT, and IEC 62443. It also warns that attackers may use unrestricted or fine tuned models while defenders face AI limits.

Microsoft AI scan uncovers decades old Linux kernel flaws

Microsoft’s FORGE Lab used AI tools under the codename MDASH to scan selected Linux kernel areas for about two weeks. The scan produced 26,502 potential findings, which researchers narrowed to 59 public bug reports after automated checks, testing with QEMU, and human review. Microsoft developed working local privilege escalation exploits for eight flaws, and 24 of the 33 assigned CVEs had High or higher severity ratings. The reports included bugs that had been in the kernel for an average of 8.1 years, with one dating back at least 21 years. At publication, fixes for 38 reports had reached the Linux mainline, while work on the rest continued.

Sources

• Attackers Use Autonomous AI to Target Semiconductor, Defense and Civil Society Sectors
• ARTEX and AI-Powered Red Teaming Tools: What Defenders Need to Know
• ‘Are robots scary?’ Phoenix Hill art exhibit examines humans’ AI thoughts, fears
• Google Pixel, Highsnobiety Showcase AI Design Products
• Why teaching 'ethical' classroom use of AI to learn writing must include the right to refuse
• Adobe Just Changed Content Creation Forever. Here's How to Stand Out in the Age of AI
• Experts: Specialized AI has great potential for regulatory, needs SME oversight
• Why training alone can't close the gap between AI adoption and ROI
• The promise and peril of AI in government
• How AI Refusals Create Blind Spots in IT/OT Security
• AI found Linux kernel bugs that remained undetected for decades