As AI agents move from pilot programs to production, experts at RSAC 2026 warn that current identity systems are ill-equipped for machine-speed actions. Matt Caulfield of Cisco and Elia Zaitsev of CrowdStrike propose a six-stage maturity model to help enterprises manage these risks, emphasizing the need for distinct identity checks to prevent agents from going rogue.
Major tech players are implementing specific safety measures to support this shift. OpenAI has released new guidelines for its Codex coding agent, utilizing sandboxing to restrict file writes and requiring human approval for high-risk actions within the ChatGPT enterprise workspace. Meanwhile, Caylent, an AWS Premier Tier Services Partner, has won a Gold Globee Award for helping businesses scale AI operations, claiming its agentic engineering services can save up to 75% of time on application modernization.
Government agencies are also leveraging AI for critical functions. The Pandemic Response Accountability Committee uses AI to shift from reactive to proactive fraud detection in federal spending, while the National Institutes of Health employs AI to break down health data silos, managing a 440 petabyte collection to link diseases like COVID and cancer. These efforts highlight a broader trend of using AI to handle vast datasets more efficiently.
Industry collaborations are accelerating innovation in specialized fields. L'Oreal Group partners with NVIDIA to use the ALCHEMI framework for skincare research, simulating ingredient performance at the atomic level to speed up discovery by up to 100 times. However, challenges remain, as Wikipedia dropped out of the top ten websites in April 2026 due to AI tools providing quick answers, while concerns about editorial bias persist.
Regulatory and security landscapes are evolving rapidly. The European Union has delayed high-risk AI system rules under the AI Omnibus deal but expanded bans on nudifier tools. Security expert Gina Scinta from Thales TCT highlights that 97 percent of organizations faced harm from AI disinformation, calling for stronger RAG system security, including pre-ingestion data discovery and post-quantum cryptography readiness.
Key Takeaways
['Cisco and CrowdStrike experts propose a six-stage maturity model to govern agentic AI identity and prevent rogue actions.', 'OpenAI deploys sandboxing and human approval requirements for its Codex coding agent within ChatGPT enterprise workspaces.', 'The Pandemic Response Accountability Committee uses AI to proactively stop fraud in federal pandemic funding.', 'The NIH utilizes AI to connect health data silos, managing a 440 petabyte collection to study disease links.', 'Caylent wins a Gold Globee Award for agentic engineering services that save up to 75% of time on application modernization.', 'The EU AI Omnibus deal delays high-risk system rules while banning nudifier tools and expanding bias detection regulations.', 'Experts suggest model tricks like fine-tuning and LoRA can slash AI training costs without buying more hardware.', 'Wikipedia fell out of the top ten websites in April 2026 as AI tools provide quick answers to users.', "L'Oreal partners with NVIDIA to use the ALCHEMI framework, accelerating skincare product discovery by 100 times.", 'Gina Scinta warns that 97% of organizations suffered from AI disinformation, urging stronger RAG system security measures.']Experts outline six stages to govern agentic AI identity
A recent interview at RSAC 2026 discusses how companies must update identity systems to manage AI agents safely. Matt Caulfield from Cisco and Elia Zaitsev from CrowdStrike explain that current tools were built for humans, not for agents that act at machine speed. They propose a six-stage maturity model to help enterprises close the gap between testing pilots and safe production use. The experts warn that agents can go rogue if they lack proper action-level controls and distinct identity checks.
OpenAI details safety controls for its Codex coding agent
OpenAI has released new guidelines on how it safely deploys its Codex coding agent in real workflows. The company uses sandboxing to limit where the agent can write files and requires human approval for high-risk actions. They also use auto-review modes to speed up routine tasks while blocking dangerous commands. Network access is strictly managed, and all credentials are stored securely within the ChatGPT enterprise workspace.
PRAC uses AI tools to stop fraud in pandemic funding
The Pandemic Response Accountability Committee is using artificial intelligence to fight fraud in federal spending. Executive Director Ken Dieffenbach says AI helps investigators analyze vast amounts of data much faster than humans could alone. The agency is shifting from a reactive model to a proactive one that stops fraudulent payments before money is released. They are also using new executive orders to verify data against other government databases like the Social Security Administration.
NIH uses AI to break down health data silos
The National Institutes of Health is using AI to connect separate health data systems for faster research. Associate Director Susan Gregurick explains that AI extracts information from pathology reports and combines it with electronic health records. This helps researchers understand links between diseases like COVID and cancer progression. The agency also uses AI to sort thousands of grant applications and manage its massive 440 petabyte data collection.
Caylent wins gold award for AI service delivery
Caylent has won a Gold Globee Award for its work as an artificial intelligence service provider. The company was recognized for helping businesses move from AI experiments to large-scale operations. As an AWS Premier Tier Services Partner, Caylent claims its agentic engineering services can save up to 75% of time on application modernization. The award highlights their focus on combining innovation with strict governance and security.
EU AI Omnibus deal delays high-risk system rules
European Union legislators have reached a deal that changes the timeline for the AI Act. The main change postpones the start date for requirements on high-risk AI systems to allow more time for preparation. The deal also expands rules on using sensitive data for bias detection and adds a ban on nudifier tools. While some obligations are simplified, the core requirements for AI safety remain largely the same.
Twelve model tricks can slash AI training costs
Experts suggest that companies can reduce AI training costs by changing how models are built rather than just buying more hardware. Key strategies include fine-tuning existing models instead of training from scratch and using parameter-efficient techniques like LoRA. Other methods involve gradient checkpointing to save memory and compiler fusion to speed up processing. These steps allow teams to achieve better results without spending millions on raw compute power.
Wikipedia drops out of top ten websites due to AI
Wikipedia has fallen out of the top ten most visited websites in April 2026, according to SimilarWeb data. The decline is attributed to AI tools that now provide quick answers without users needing to visit an encyclopedia site. Google traffic has grown during this same period, showing resilience against AI disruption. Critics also point to concerns about Wikipedia's editorial bias as another reason users are leaving the platform.
L'Oreal and NVIDIA speed up beauty product research
L'Oreal Group is partnering with NVIDIA to use AI for developing new skincare products. The collaboration uses the NVIDIA ALCHEMI framework to simulate ingredient performance at the atomic level. This digital approach allows scientists to test thousands of variables virtually, which accelerates the discovery process by up to 100 times. The goal is to create more effective products for skin protection and aging prevention faster than traditional lab methods.
AI impacts democracy and family privacy in 2026
An article discusses how artificial intelligence is affecting democracy and everyday life as the 2026 midterm election approaches. Senator Bernie Sanders is quoted discussing the risks of profit-driven data collection and political targeting by AI companies. The piece includes interviews with different AI models like Grok, ChatGPT, and DeepSeek about their views on democracy. The conversation highlights tensions between truth-seeking and safety controls in AI systems.
Gina Scinta calls for stronger security in RAG systems
Gina Scinta from Thales TCT warns that RAG systems create new security risks for sensitive data. She notes that 97 percent of organizations surveyed experienced harm from AI-generated disinformation. To fix this, she recommends that agencies require vendors to offer pre-ingestion data discovery and transparent encryption. Other necessary capabilities include independent key management and post-quantum cryptography readiness to protect data from future threats.
Sources
- AI agent identity: how to govern agentic AI in 6 stages
- Running Codex safely at OpenAI
- AI & Data Exchange 2026: PRAC’s Ken Dieffenbach on using AI tools to stay a step ahead of fraudsters
- AI & Data Exchange 2026: NIH’s Susan Gregurick on overcoming data silos with AI analytics
- Caylent Wins Gold at the 2nd Annual 2026 Globee® Awards for Artificial Intelligence
- What the EU AI Omnibus Deal Changes for the AI Act and What Lies Ahead
- 12 model-level deep cuts to slash AI training costs
- AI Impact? Wikipedia Drops Out Of The Top 10 Websites By Traffic List In April
- 100 times faster: L’Oréal and NVIDIA are transforming the beauty industry
- AI, Democracy And The Politics Of The Kitchen Table
- Gina Scinta Urges Stronger AI Security for RAG Systems
Comments
Please log in to post a comment.