Nvidia joins AI security summit as Anthropic launches Claude Mythos

Key Takeaways

• Security researchers identified six critical AI vulnerabilities between mid-2025 and April 2026 affecting enterprise platforms.
• Mend.io released a risk tier system for AI security, with high-risk models scoring 12-15 requiring full assessments.
• Singapore issued new governance guidelines specifically for agentic AI systems and identity management.
• SCAD launched an Applied AI degree featuring a curriculum with NVIDIA, Google, and Adobe leaders.
• IBM shareholders voted on Proposal 7 to audit potential bias in the company's AI models.
• A former Google employee was convicted of stealing AI secrets to provide to Chinese firms.
• India is shifting to stricter AI regulations following controversies involving Grok and Anthropic's Claude Mythos.
• A new brain-like chip from Cambridge University could reduce AI energy consumption by up to 70%.
• The Linux kernel community is debating removing legacy network drivers to stop AI-generated bug reports.
• Penn State Berks hosted a free workshop for K-12 educators on designing critical thinking assessments in the age of AI.

Six Critical AI Vulnerabilities Expose Major Security Gaps

Security researchers identified six critical AI vulnerabilities between mid-2025 and April 2026 that affect platforms used by many enterprises. These flaws follow three distinct attack patterns involving untrusted input, overly broad data access, and process containment failures. The most common issue is that external data like emails or documents is processed by AI without proper validation. Five of the six vulnerabilities involved AI systems accessing far more data than intended due to lack of per-operation access controls. The remaining case involves architectural failures in how GrafanaGhost handled event monitoring logs.

Mend.io Launches New AI Security Governance Framework

Mend.io released a practical framework guide to help organizations manage AI security risks effectively. The guide introduces a risk tier system where low-risk models score 5-7, medium-risk models score 8-11, and high-risk models score 12-15. High-risk tiers require full security assessments and incident response playbooks. The framework also mandates output filtering for sensitive data like SSNs and credit card numbers. It emphasizes that AI-generated code must undergo the same security scans as human-written code. The guide defines three monitoring layers to catch threats that traditional SIEM tools often miss.

Linux Kernel May Remove Old Network Drivers Due to AI Reports

The Linux kernel community is debating a proposal to remove legacy network drivers from the main source code. This change aims to stop an unsustainable surge in bug reports generated by AI tools and fuzzers targeting unused old hardware. Andrew Lunn argues that maintaining support for ISA and PCMCIA-era devices has become a disproportionate burden. If accepted, the proposal would eliminate approximately 27,646 lines of code covering hardware from companies like 3Com, AMD, and Cirrus Logic. The removal would happen one patch at a time to allow users to restore drivers if they still need them.

Wiser Shares Thoughts on Words, Grammar, and AI

Chuck Wiser wrote a weekly column discussing language quirks, grammar issues, and the influence of AI on word usage. He explored the word Florificent, noting it does not appear in standard dictionaries but seems to be a creative compound of Flor and -ificent. Wiser also examined the pronunciation of the 'ph' sound in words like phonograph and compared it to the 'gh' sound. He mentioned April snow flurries and his efforts to rebuild a sunflower seed bird feeder for Evening Grosbeaks and Baltimore Orioles. The piece concluded with a report on a poetry reading event at the David A. Howe Library.

Singapore Issues New Guidelines for Agentic AI Security

Singapore released governance and security guidance specifically designed for agentic AI systems. Agentic AI refers to systems that can plan multiple steps, take actions, and interact with external systems to achieve goals. The framework highlights risks such as hallucinated plans, tool misuse, and cascading errors in multi-agent systems. It recommends that organizations assess risks upfront by considering the domain, data access, and level of autonomy. The guidelines also stress the importance of agent identity management and ensuring humans remain accountable for agent actions.

Penn State Berks Hosts AI Assessment Workshop for Teachers

Penn State Berks will host a workshop titled Designing Critical Thinking Assessments in the Age of AI for K-12 educators. The event takes place on May 12 from 8:30 a.m. to 2 p.m. in the Gaige Technology and Business Innovation Building. Participants will receive five hours of ACT 48 credit and the event is free with meals provided. The workshop is supported by the U.S. National Science Foundation through the SMILE project. It teaches educators how to use prompt engineering to turn large language models into reliable teaching assistants. The afternoon session focuses on designing assessments that build critical thinking through Socratic questioning.

Critics Demand IBM AI Model Bias Audit at Annual Meeting

Shareholders of IBM will vote on a proposal from the National Center for Public Policy Research to audit potential bias in the company's AI models. The proposal, called Proposal 7, requests a report on data sources, bias mitigation methods, and output accuracy. Steve Milloy, an AI user, argued that AI models often cannot distinguish between good information and garbage found on the internet. He cited the climate change debate as an example where AI spreads misinformation learned from erroneous online content. Milloy claimed IBM's own website contributes to this pollution by promoting claims about global warming that he considers false.

Former Google Employee Convicted of Stealing AI Secrets for China

A former Google employee named Linwei Ding was convicted of stealing AI secrets to give to Chinese firms. He downloaded sensitive data and sought to use the stolen technology to build AI systems in China. Tom Lyons testified before the Senate Judiciary Committee that Chinese espionage represents a national security threat rather than normal competition. Lyons warned that American companies are competing against the largest intelligence apparatus in the world. He argued that the current approach leaves firms largely on their own to counter state-backed threats. Chinese officials have repeatedly denied engaging in such activity.

SCAD Introduces Applied AI Degree for Creative Professionals

Savannah College of Art and Design launched a new Applied AI degree to prepare students for careers in creative technology. The program teaches students to design intelligent systems that balance innovation with empathy and responsibility. Students will learn to build interactive agents, immersive environments, and intelligent products that connect with users. The curriculum includes ethics, behavior, and cultural impact considerations. SCAD hosts an annual AI Summit bringing together leaders from companies like NVIDIA, Google, and Adobe. Students also participated in a hands-on AI Jam powered by NVIDIA to prototype generative tools.

India Considers Stricter AI Regulations After Content Controversies

India is moving away from light-touch AI regulations toward a more prescriptive approach due to rapid technological advances. A six-member committee called TPEC and a 10-member inter-ministerial group are working on new guidelines. Recent controversies involving explicit AI-generated content on Grok and the launch of Anthropic's Claude Mythos model have raised concerns. Officials noted that these developments could put critical sectors like finance and energy at risk. The groups aim to create a tri-model approach that addresses current challenges and prepares for future risks.

New Brain-Like Chip Could Reduce AI Energy Use by 70%

Researchers at the University of Cambridge developed a new nanoelectronic device that could slash AI energy consumption by up to 70%. The device uses a modified hafnium oxide material to create stable, low-energy memristors that mimic how neurons connect. Unlike traditional chips that separate memory and processing, this neuromorphic approach combines them to reduce data transfer needs. The new design avoids unpredictable conductive filaments by using controlled energy barriers at interfaces. Tests showed the devices operate at currents a million times lower than conventional memristors and can learn and adapt like biological neurons.

Sources

• Six AI Vulnerabilities, Three Attack Patterns, One Dangerous Service Gap
• Mend.io Releases AI Security Governance Framework Covering Asset Inventory, Risk Tiering, AI Supply Chain Security, and Maturity Model
• Linux may be ending support for older network drivers due to influx of false AI-generated bug reports — maintenance has become too burdensome for old largely-unused systems
• Wiser’s Wramblings…“Florificent,” and Phooled AI, As Did the F Word(s)
• Singapore Issues Governance and Security Guidance for Agentic AI
• Penn State Berks to host AI-focused assessment workshop for K-12 educators
• IBM’s AI Model: Garbage In, Garbage Out
• Google engineer stole AI secrets for China, Senate hears in explosive testimony
• Applied AI at Savannah College of Art and Design
• AI’s rapid rise forces India to rethink ‘light-touch’ regulations
• This new brain-like chip could slash AI energy use by 70%