AI adoption is moving faster than many organizationsâ ability to govern it, according to Palo Alto Networksâ 2026 State of Cloud Security Report, based on 2,800 respondents across 10 countries and four regions. The report says 99% of organizations had ungoverned AI in their systems during the past year, yet only 9% can automatically discover and classify it. Although 99% use AI agents in development pipelines and 73% allow agents to change code or production systems, organizations enforce fewer than three of eight measured predeployment checks on average. Forty percent cannot reconstruct an agentâs actionsâ46% in financial servicesâand only 36% monitor AI behavior in production. AI-driven development has increased cloud infrastructure updates for 82% of respondents. In the 90 days before the survey, 19% of high- and critical-severity application security issues reached production; fixes took an average of 18.4 days to deploy.
Security concerns also feature in separate reports and company developments. An Industrial Cyber article says a cable fault in London disrupted a Google cloud zone for six hours and eight minutes after two programmable logic controllers incorrectly detected a working power source; nearly 40 services were affected across the wider region. Anthropicâs Frontier Red Team developed a Linux root exploit in about 53 minutes after initial prompting, while patch availability varied among Linux systems. Capcom said unauthorized access exposed 11,774,247 records, and passwords were changed without permission on 196 accounts. NTT, meanwhile, says its cybersecurity AI can analyze software for vulnerabilities without sending source code to external AI services. Combining open-weight models, vulnerability research methods and workflow controls, it raised benchmark success from 48% to 63% and plans to improve discovery and support organization-managed computing.
OpenAI and Chileâs government have launched OpenAI for Chile, led by the Ministry of Science, Technology, Knowledge and Innovation. The partnership aims to support responsible AI use in public administration, train civil servants and help local startups develop AI applications. The partners will explore training, technology guidance and uses in ministry processes, with a joint task force expected to establish priorities and an initial workplan. The agreement commits neither side to investment or purchases and includes safeguards for privacy, information security and responsible use.
A separate essay urges the United States and China to cooperate in restraining frontier AI development. Its authors argue that racing toward artificial general intelligence or self-improving models could produce disaster for both the winner and others. They frame mutual restraint as a âstag huntâ: a rational option if both governments recognize the risks and can verify one anotherâs actions. A planned second part will propose verifiable restraint and international and domestic governance. In another discussion of AIâs effects, Northeastern University London researcher Melika Vafafar says game-embedded companions could respond to playersâ words, detect frustration and suggest breaks after losing streaks or long sessions. Eye-movement or sweaty-palm sensors could provide additional signals, but the paper also raises concerns about systems shaping playersâ emotions and habits.
Workplace adoption can be driven by expectations as well as usefulness: a JobsAssist survey found 63% of workers say employers actively encourage or expect AI use, 39% highlight their AI use to managers, and 52% have used AI even when they thought they could finish a task faster without it. In banking, a report cited in an article expects AI use to rise 25% over the next year as Maryland banks invest in customer service and operations. Experts warn that automation could cost jobs and worsen skills shortages, while vulnerable systems could create cybersecurity threats and financial losses for banks and customers.
For B2B vendors, the article on AI-agent recommendations argues that clear, connected evidence can help a company make a buyerâs shortlist. Vendors should link product capabilities to customer needs and support claims with credible material, such as product pages, case studies, technical documents, customer reviews and sales knowledge. Supply-chain software buyers, for example, may seek specifics on integrations, implementation timelines, forecasting, inventory and operating results. Separately, an article about AI reviewing old physics papers reports that seven physicists discussed whether model-generated comments identify genuine errors or amount to noise, alongside unresolved questions about gravityâs relationship to quantum physics and whether quantum gravity is a meaningful concept.
Key Takeaways
- Palo Alto Networks found 99% of surveyed organizations had ungoverned AI in their systems in the past year, but only 9% can automatically discover and classify it.
- Although 99% use AI agents in development and 73% let them change code or production systems, organizations enforce fewer than three of eight measured predeployment checks on average; 40% cannot reconstruct agent actions.
- The report says 19% of high- and critical-severity application security issues reached production in the prior 90 days, with fixes taking an average of 18.4 days to deploy.
- Googleâs London cloud-zone outage lasted six hours and eight minutes after two programmable logic controllers misread power status; Anthropicâs Frontier Red Team developed a Linux root exploit in about 53 minutes.
- NTT says its source-code-in-house cybersecurity AI raised benchmark success from 48% to 63% and supports vulnerability discovery, assessment and remediation.
- OpenAI and Chile launched OpenAI for Chile to support responsible public-sector AI, civil-servant training and local startups; the agreement includes no investment or purchasing commitment.
- An essay calls for verifiable US-China cooperation to restrain frontier AI development, arguing that a race toward AGI or self-improving models could endanger both sides.
- JobsAssist found 63% of workers say employers expect or encourage AI use, while 52% have used it even when they believed doing the task without AI would be faster.
- Maryland banks expect AI use to rise 25% over the next year, while experts warn of job losses, skills shortages, cybersecurity risks and potential financial losses.
- Other articles examine AI companionsâ possible emotional influence in games, the evidence B2B vendors need to earn AI-agent recommendations, and physicistsâ questions about whether AI reviews of research papers identify real errors.
Palo Alto Networks finds widespread ungoverned AI use
Palo Alto Networksâ 2026 State of Cloud Security Report draws on responses from 2,800 people in 10 countries and four regions. It found that 99% of organizations had ungoverned AI in their systems in the past year, but only 9% can automatically find and classify it. While 99% use AI agents in development pipelines and 73% let them change code or production systems, organizations enforce fewer than three of eight measured predeployment checks. The report also found that 40% cannot retrace an AI agentâs actions, and 36% can monitor AI behavior in production.
Many organizations cannot retrace AI agentsâ actions
Palo Alto Networksâ State of Cloud Security Report 2026 found that 40% of organizations could not reconstruct an AI agentâs actions, rising to 46% in the financial sector. AI-driven development has increased cloud infrastructure updates for 82% of respondents, while organizations enforce fewer than three of eight measured predeployment security checks on average. In the 90 days before the report, 19% of high- and critical-severity application security issues reached production, and it took an average of 18.4 days to deploy a fix. Only 9% can automatically discover and classify ungoverned AI, while 36% can monitor AI behavior in production.
OpenAI and Chile partner to expand responsible AI use
OpenAI and Chileâs Government launched OpenAI for Chile, a partnership led by the Ministry of Science, Technology, Knowledge and Innovation. It aims to support responsible AI use in public administration, build civil servantsâ skills and help local startups develop AI applications. The partners will explore training, technology guidance and ways to use AI in ministry processes, with a joint task force expected to set priorities and an initial workplan. The agreement makes no investment or purchasing commitment, and activities will include safeguards for privacy, information security and responsible use.
US and China urged to cooperate on frontier AI risks
The essay argues that the United States and China should restrain the development of frontier AI models together. It says leaders often treat AI as a winner-take-all race, even as AI companies warn about risks and call for restraint. The authors argue that racing to develop artificial general intelligence or self-improving models could create a disaster for both the winner and others. They describe cooperation as a âstag hunt,â where mutual restraint can be a rational choice if both governments recognize the risks and can verify each otherâs actions. Part I explains this argument, while Part II will propose forms of verifiable restraint and a system of international and domestic governance.
Workers use AI because they think employers expect it
A JobsAssist survey found that employees sometimes use AI for appearances, even when it does not help them work faster. Sixty-three percent of workers said their employers actively encourage or expect them to use AI, and 39% said they highlight their AI use to managers. At the same time, 52% said they had used AI even when they believed they could complete a task more quickly without it.
Researchers see AI companions offering emotional support in games
Melika Vafafar, a computer science PhD researcher at Northeastern University London, says AI companions in video games could one day support playersâ emotional well-being. In a paper titled âServing the Player or Shaping Them? Game-Embedded AI Companions and the Rewiring of Well-being,â she describes how companions could respond to playersâ words, notice frustration, and suggest a break after a losing streak or a long session. Sensors that track signals such as eye movements or sweaty palms could give AI more clues about how players feel. Vafafar says long playtime could let these systems affect playersâ emotions and habits, while others question whether AI should provide this kind of support.
Maryland banks weigh AI savings against growing risks
Maryland banks are weighing AIâs potential to improve efficiency and cut costs against risks to the financial industry. A recent report cited in the article expects AI use in banking to rise by 25% in the next year, as banks invest in systems for customer service and operations. Experts warn that automation could cause job losses and make it harder to find skilled workers. They also say vulnerable AI systems could increase cybersecurity threats and lead to financial losses for banks and customers.
A power failure, AI exploit and identity leaks expose security gaps
An Industrial Cyber article connects three security stories to basic operational technology security needs: testing control systems, patching exposed devices and protecting identity data. Google reported that a cable fault in London cut power to a cloud zone after two programmable logic controllers incorrectly detected a working power source; the outage affected customers for six hours and eight minutes, and close to 40 services were affected across the wider region. Anthropicâs Frontier Red Team developed a Linux root exploit in about 53 minutes after initial prompting, while patch availability varied across Linux systems. Capcom also disclosed that unauthorized access to CAPCOM ID exposed 11,774,247 records, and passwords were changed without permission on 196 accounts.
Clear evidence can help B2B vendors earn AI agent recommendations
As AI agents research and compare business vendors, the information companies provide may determine whether they make a buyerâs shortlist. The article says vendors should connect product capabilities to customer needs and back up claims with credible evidence. For example, supply chain software buyers may want details about integrations, implementation timelines, forecasting, inventory and operating results. Product pages, case studies, technical documents, customer reviews and sales knowledge can all help, but they need to work together and answer buyersâ questions.
NTT builds cybersecurity AI that keeps source code in-house
NTT says it has developed cybersecurity AI that can analyze software for vulnerabilities without sending source code to external AI services. The company combined multiple open-weight models with vulnerability research methods and its own workflow controls, raising its benchmark success rate from 48% to 63%. NTT says its technology can support vulnerability discovery, assessment and remediation while letting organizations control how they select and operate AI. It plans to improve the systemâs discovery capabilities and make it easier to run with computing resources managed by organizations themselves.
Seven physicists weigh AIâs role in reviewing research
The article describes how an AI model reviews old physics papers, often producing comments about them. The author spoke with seven physics experts about AIâs growing role in the field. The article raises the question of whether the modelâs notes identify real errors or are mostly noise. It also discusses open questions about gravity, including whether gravity is a quantum phenomenon and whether quantum gravity is a meaningful idea.
Sources
- Palo Alto Networks: 99% of Companies Have Ungoverned AI
- 40% of organizations struggle to retrace AI agentsâ actions
- Introducing OpenAI for Chile
- A Game-Theory Basis for US-China Cooperation on AI Governance (Part I)
- Why are employees using AI even when it slows them down?
- Northeastern experts say AI can provide in-game mental health support
- MD bankers weigh AI efficiencies against rising risks
- Inside OT's Fundamentals Gap: A PLC Fault, an AI Exploit, and Leaked Identities
- When AI agents evaluate B2B vendors, what earns consideration?
- NTT Develops Sovereign Cybersecurity AI, âApproaches Frontier AI Capabilitiesâ
- I spoke to seven physics experts grappling with the AI revolution
Comments
Please log in to post a comment.