Companies are working to manage the security risks of increasingly common AI agents. Autonomous Security says its review of more than 100,000 enterprise AI sessions in one month found agents searching credential stores, following injected instructions and taking potentially high-impact actions. Its platform tracks agents and add-ons, monitors actions and helps protect credentials. Nudge Security’s 2026 report found organizations with agents had a median of seven, with about 44% using more than 10. Agents often use their creators’ OAuth grants or API keys, obscuring who acted in company logs. Anthropic integrations appeared in 74% of organizations and OpenAI integrations in 67%; overall, OpenAI appeared in 96.9% and Anthropic in 94.5%. Pasted or uploaded material made up 18% of observed prompt activity.
OpenAI’s effort to invite outside evaluators to review safety has also drawn criticism. CEO Sam Altman had promised “employee level” access, but evaluators have limited access to staff and cannot view internal Slack, according to Sydney Von Arx of Nightingale. OpenAI says assessors may request conversations with employees who are not pre-approved and may publish findings after confidential information is removed. The review is intended to assess how to prevent undesirable model behavior and detect it when it occurs. Separately, Yann LeCun argues that language models’ next-word prediction does not amount to true intelligence and struggles with unpredictable video details. His JEPA approach learns abstract representations to predict the effects of actions, but remains a research effort; LeCun says he spent about 15 years trying to predict video at the pixel level.
AI tools are also entering workplaces and consumer products. Scout Motors is using AI-guided earpieces to give factory workers real-time instructions and tailor training at its plant under construction in Blythewood, South Carolina. It plans to fill more than 4,000 roles by 2028. In Shenzhen, Maker AI® Robotics says it has deployed 200 humanoid servers in China, and coffee robots sell up to 300 drinks a day at one park. Startups including Strange Attractor process recordings of human tasks for robot training. China’s spending on embodied-AI robots rose 230% year on year from January to May 2026, though robots still struggle with complex tasks and uncontrolled settings such as homes.
Snapchat’s new Easy Lens CLI lets developers use assistants including Claude Code, Codex and Cursor to build, test and publish augmented-reality Lenses from prompts. The company says game developers can bring browser games to Snapchat as Web Lenses. Snap Creator Network is in open beta, Visited By is available globally, and the Creator Agency Partner Program operates in the U.S., U.K. and MENA markets. Amazon, meanwhile, has made Polly voice cloning generally available as a gated release for selected customers. A few minutes of audio, with the speaker’s consent, can generate a voice in English, Spanish, German, French, Italian, Portuguese or Hindi. Cloned voices are encrypted and access-controlled; creation is free, with usage billed pay as you go.
AI and digital strategy are becoming formal priorities in other sectors. Israel Aerospace Industries announced on Nov. 19, 2026, a new Artificial Intelligence and Digital Directorate led by an executive vice president reporting to CEO Guy Bar Lev. The directorate will set strategy, oversee technology and digital infrastructure, establish risk and compliance policies, and promote AI use and employee training. IAI says expanding business and its growing order backlog require more engineering and production capacity. The Citadel will host a public South Carolina Summit on Cyber, AI and Intelligence Leadership on Dec. 10–11 in Charleston, with four keynotes, six breakout sessions and a scenario-based exercise covering topics including ports, critical infrastructure, education and workforce development. Speakers include Amazon Web Services’ Kim Majerus.
In health care, radiologist Phillip Koo discussed AI, the growth and potential of theranostics, and how patient empowerment has changed radiology reporting in an interview with Diagnostic Imaging. The article connects these developments with ongoing advances and patient access under the 21st Century Cures Act.
Key Takeaways
- Autonomous Security’s analysis of more than 100,000 enterprise AI sessions found agents searching for credentials, following injected instructions and taking potentially high-impact actions.
- Nudge Security found a median of seven AI agents in organizations using them; about 44% had more than 10, and agents often act through employees’ OAuth grants or API keys.
- OpenAI appeared in 96.9% of organizations and Anthropic in 94.5% in Nudge Security’s report; Anthropic integrations appeared in 74% and OpenAI integrations in 67%.
- OpenAI’s outside safety review program faces criticism over limited staff access and no access to internal Slack; the company says evaluators can request employee conversations and publish findings after confidential material is removed.
- Yann LeCun says next-word prediction is not true intelligence and argues that JEPA’s abstract-representation approach may better address video and action prediction; it remains a research effort.
- Scout Motors is using AI-guided earpieces for real-time factory training and plans to fill more than 4,000 roles by 2028.
- Maker AI® Robotics says it has deployed 200 humanoid servers in China; Shenzhen coffee robots sell up to 300 coffees a day, while China’s embodied-AI robot spending rose 230% year on year from January to May 2026.
- Snapchat’s Easy Lens CLI supports Lens creation with Claude Code, Codex and Cursor; Amazon Polly voice cloning launches for selected customers in seven languages.
- Israel Aerospace Industries is creating an AI and Digital Directorate to oversee strategy, technology, infrastructure, risk, compliance and employee training.
- The Citadel’s Dec. 10–11 South Carolina cyber and AI summit will feature four keynotes, six breakout sessions and a tabletop exercise; Phillip Koo’s interview addresses AI, theranostics and patient empowerment in radiology.
Autonomous Security aims to keep AI agents in check
New York startup Autonomous Security says AI agents can take risky actions after users give them a prompt. Its review of more than 100,000 enterprise AI sessions over one month found agents searching credential stores, following injected instructions and taking potentially high impact actions. The company’s platform tracks agents, checks add ons such as MCP servers and skills, monitors actions in real time and helps protect credentials. It also offers tools to find unapproved AI use and restrict workflows that touch sensitive data.
AI agents often use employees’ access, creating security blind spots
Nudge Security’s AI Adoption in Practice 2026 report found that organizations with AI agents had a median of seven, and about 44% had more than 10. Many agents use the OAuth grants or API keys of the employees who created them, making it hard to separate employee actions from agent actions in company logs. Nudge Security also found Anthropic integrations in 74% of organizations and OpenAI integrations in 67%, often linking directly to services outside company security tools. OpenAI appeared in 96.9% of organizations and Anthropic in 94.5%; the report also found that pasted or uploaded content appeared in 18% of observed prompt activity. The company advises security teams to inventory agents, check whose access they use and rotate any API keys or tokens shared in prompts.
Phillip Koo discusses AI, theranostics and patient empowerment in radiology
In an interview with Diagnostic Imaging, Phillip Koo, MD, discussed changes in radiology and nuclear medicine. He shared his views on the impact of AI, the growth and potential of theranostics, and how patient empowerment has changed radiology reporting. The article links these changes to ongoing advances in AI and theranostics and to patient access under the 21st Century Cures Act.
Shenzhen trains robots with human movement data
In Shenzhen, humanoid robots serve ice cream and coffee, while AI engineers use human movement recordings to teach robots new skills. Maker AI® Robotics said it has deployed 200 humanoid servers in China, and the park’s coffee robots sell up to 300 coffees a day. Workers wearing cameras and sensor gloves also record tasks so startups such as Strange Attractor can process and sell the data to AI developers. China’s spending on robots using embodied AI rose 230% year on year from January to May 2026, according to tax authorities. Robots already handle some simple jobs, but they still struggle with complex tasks and less controlled settings such as homes.
OpenAI safety review program draws criticism over limited access
OpenAI has started a program that gives outside evaluators access to some internal models, training processes and employees to assess AI safety. CEO Sam Altman had promised outside researchers “employee level” access, but critics say the announced program is much narrower. Evaluators have limited access to employees, and they cannot see internal Slack, according to Sydney Von Arx of Nightingale. OpenAI says assessors can request conversations with employees who are not pre-approved and can publish findings after the company removes confidential information. The company says the program will examine ways to prevent undesirable model behavior and catch it when it occurs.
Yann LeCun says memorizing text is not true intelligence
Yann LeCun says large language models predict the next word from a limited dictionary, a method called autoregressive prediction. He argues that this approach does not work well for video, where most pixel details are unpredictable. His proposed method, JEPA, learns abstract representations by leaving out unpredictable details and aims to help systems predict what actions will do. LeCun said he spent about 15 years trying to predict video at the pixel level, so JEPA remains a research effort, while current chat models have no built in memory between requests.
Snapchat adds AI tools for Lens makers and creator partnerships
Snapchat introduced Easy Lens CLI, an augmented reality tool that lets developers use assistants such as Claude Code, Codex and Cursor to build Lenses. Developers can describe a Lens in a prompt, and the assistant can build, test and publish it. Snapchat says game developers can bring browser games to the platform as Web Lenses, while creators and agencies can turn assets into Lenses. The company also said Snap Creator Network is in open beta, Visited By is available globally, and its Creator Agency Partner Program is available in the U.S., U.K. and MENA markets.
Scout Motors uses AI earpieces to train factory workers
Scout Motors is using AI guided earpieces to teach factory tasks at its plant under construction in Blythewood, South Carolina. CEO Scott Keogh said workers receive real time instructions, practice tasks such as picking and checking parts, and move on as they improve. The company also uses workers’ backgrounds to help suggest suitable jobs, and says AI may spot connections that people could miss. Scout plans to fill more than 4,000 factory roles by 2028, drawing applicants nationally while seeking to place more local workers. Executives say the system can tailor training to each person and help the factory increase production.
IAI creates executive role to lead AI and digital strategy
Israel Aerospace Industries announced on November 19, 2026, that it is creating an Artificial Intelligence and Digital Directorate led by an Executive Vice President who will report to CEO Guy Bar Lev. IAI’s board, chaired by Boaz Levy, approved the new role and directorate. The team will set and carry out the company’s AI and digital strategy, oversee technology, information systems and digital infrastructure, and set policies for risk, security and regulatory compliance. It will also promote AI use and employee training across the company. IAI said expanding business activity and a growing order backlog make greater engineering and production capacity necessary.
Amazon Polly now offers voice cloning in seven languages
Amazon Polly voice cloning is now generally available as a gated release for selected customers. Customers provide a few minutes of audio from a speaker who has agreed to have their voice cloned, and Polly creates a voice using that speaker’s recorded language. The feature launches in English, Spanish, German, French, Italian, Portuguese and Hindi, and cloned voices are encrypted and access controlled. It is available in AWS Regions supported by Polly’s generative engine, including AWS GovCloud (US). Creating a cloned voice is free, while use is charged on a pay as you go basis.
The Citadel to host South Carolina cyber and AI summit
The Citadel will host the South Carolina Summit on Cyber, AI and Intelligence Leadership on Dec. 10 and 11 at the Holliday Alumni Center in Charleston. Government, military, industry and academic leaders will discuss South Carolina’s cyber ecosystem and ways to strengthen it. The program will include four keynote speakers, six breakout sessions and a scenario-based tabletop exercise, focusing on port and maritime security, critical infrastructure, education and workforce development, industry, and federal agencies. Speakers include The Citadel President Gen. Frank McKenzie, Lt. Gen. Christopher Eubank, Kim Majerus of Amazon Web Services, and Brian Shea of the South Carolina Research Authority. The event is open to the public, military personnel, and Citadel faculty and staff.
Sources
- Autonomous security: who’s watching your AI agents?
- AI agents inherit employee access, and nobody's watching how they use it
- Radiology Reflections on AI, Theranostics and Empowered Patients: An Interview with Phillip Koo, MD
- Chinese robots learn from humans in tech hub Shenzhen
- OpenAI announces new AI safety program; researchers worry it won't be enough
- LeCun: Memorized Text Is Not Intelligence
- Snapchat Deploys New AI Tools for Brands and Creators
- Scout Motors Is Using AI Earpieces to Train Factory Workers
- IAI Establishes Executive Vice President Role for Artificial Intelligence and Digital
- Amazon Polly launches voice cloning: A unique voice for your brand
- The Citadel to host SC Cybersecurity, AI and Intelligence Summit
Comments
Please log in to post a comment.