Nvidia is taking a firm stance on AI safety by launching the Open Agent Safety Platform, a tool designed to help companies monitor every action taken by AI agents and enforce safety policies. The platform was developed with over 100 industry partners and comes on the heels of incidents where AI systems accessed files without permission. CEO Jensen Huang framed safety as essential to unlocking AI's full potential. Alongside these announcements, Nvidia revealed a record $150 billion stock buyback program.
In a separate but related development, Huang also advocated for US-China cooperation on managing AI risks, arguing that competitors can still collaborate on technology safety. He drew a comparison between AI safety and the gradual improvement of car safety over time. His comments raise the possibility that China may once again allow its companies to purchase Nvidia chips.
Meanwhile, the AI industry is buzzing ahead of a White House summit that will bring together top tech leaders, including Huang, Anthropic's CEO, and others. The event aims to balance innovation with oversight. Anthropic is reportedly seeking a valuation exceeding $2 trillion in its upcoming IPO, which would surpass SpaceX's record. German AI stocks rallied in anticipation, with companies like Infineon and Siemens Energy gaining up to 5 percent.
On the security front, researchers have introduced a method called formation-consistent dispatch (FCD) to address a vulnerability known as schema-epoch drift in AI agents. This issue can cause valid tool calls to change their security effects during execution. Testing across 32 official releases showed that 29 required no changes, suggesting the approach is both practical and effective.
Concerns about AI security were amplified by a report on Zhipu AI's GLM-5.3 model. Researchers found that attackers can bypass the model's safeguards between 64 percent and 100 percent of the time. Unlike Claude models, GLM-5.3 is openly available and lacks limits on misuse. NIST's CAISI highlighted these risks in a September 17 report, noting that GLM-5.3 successfully identified unknown browser vulnerabilities and created working exploits in tests.
Bank of America issued a warning that AI agents like Meta's Muse could threaten Apple's dominance in smartphone transactions. App Store purchases and similar revenue streams could be disrupted if AI agents make it easier for users to switch to Samsung or Google phones. The report suggests Apple's grip on the smartphone market faces real risk from this shift.
On a more commercial note, KIA Italia has found success using an AI-powered WhatsApp sales agent developed with Infobip. The campaign connected Facebook and Instagram ads to a conversational AI tool that answered customer questions about models, prices, and financing. Over two months, the system generated 144 qualified leads, 116 quote requests, and 8 direct car sales, while cutting the cost per vehicle sold by 53 percent compared to traditional campaigns.
Elsewhere, Voorhees University in South Carolina received a $1,188,800 grant from the U.S. Department of Labor to launch an AI Institute of Technology and Research Center. The funding will integrate AI skills into academic programs and expand workforce training for surrounding rural communities through September 2031. Enable Ventures also closed its first fund at $50.31 million, dedicated to AI startups that help people with disabilities in areas like worktech and digital health.
Key Takeaways
- <ul><li>Nvidia launched the Open Agent Safety Platform with 100+ partners to monitor AI agent actions and enforce safety policies.</li><li>Nvidia also announced a record $150 billion stock buyback program.</li><li>CEO Jensen Huang called for US-China cooperation on AI safety, comparing it to improvements in car safety over time.</li><li>Anthropic is seeking a $2 trillion+ valuation for its upcoming IPO, potentially surpassing SpaceX's record.</li><li>Bank of America warns that Meta's Muse AI agent could weaken Apple's iPhone transaction dominance and App Store revenue.</li><li>Researchers developed formation-consistent dispatch (FCD) to fix schema-epoch drift in AI agents, tested across 32 releases with positive results.</li><li>Zhipu AI's GLM-5.3 model has safeguard bypass rates between 64% and 100%, according to NIST's CAISI report.</li><li>KIA Italia's AI WhatsApp agent generated 8 direct car sales and cut cost per vehicle sold by 53% in two months.</li><li>Voorhees University received a $1.19M grant to launch an AI Institute running through September 2031.</li><li>Enable Ventures closed a $50.31M fund focused on AI startups serving people with disabilities.</li></ul>
Nvidia launches platform to prevent AI agents from going rogue
Nvidia introduced the NVIDIA Open Agent Safety Platform with over 100 industry partners. The tool lets companies monitor every action by an AI agent and enforce safety policies. This comes after incidents where AI systems accessed files without permission. Nvidia CEO Jensen Huang called AI safety essential for realizing AI's benefits. Nvidia also announced a $150 billion share buyback, the largest ever.
Jensen Huang discusses US-China AI cooperation and safety
Nvidia CEO Jensen Huang said the US and China should collaborate to manage AI risks. He stated that competitors can still work together on technology. Huang discussed China potentially allowing its companies to buy Nvidia chips again. He compared AI safety to car safety improvements over time. Nvidia also announced a record $150 billion stock buyback alongside these remarks.
KIA Italia achieves 40x return on ad spend using AI WhatsApp sales
KIA Italia partnered with Infobip to create an AI-powered WhatsApp sales experience. The campaign used Facebook and Instagram ads that connected users to an AI agent. In two months, it generated 144 qualified leads, 116 quote requests, and 8 direct car sales. The cost per vehicle sold dropped 53% compared to traditional campaigns. The AI agent answered questions about models, prices, and financing in real time.
KIA Italia uses AI WhatsApp agent to boost car sales
KIA Italia worked with Infobip to build a conversational sales tool on WhatsApp. The system integrated with Salesforce Marketing Cloud and CRM to manage leads automatically. Customers could interact with the AI agent for instant answers about car models and financing. The campaign averaged over 10 messages per session. Infobip's Ben Lewis and KIA Italia's Giuseppe Mazzara praised the results as a measurable sales channel.
New method addresses security drift in AI language model agents
Researchers presented formation-consistent dispatch (FCD) to fix a problem called schema-epoch drift in AI agents. This issue occurs when valid tool calls change security effects during rollout or delayed approval. FCD connects implementation analysis to execution authority using verified profiles. Testing across 32 official releases showed 29 required no changes. The method was tested in stock comparisons and lifecycle experiments with positive results.
AI stocks rally as White House AI summit and Anthropic IPO approach
German AI-related stocks rebounded on Tuesday after a weak start to the week. Shares of Infineon, Siemens Energy, and other tech companies rose by up to 5 percent. Attention is shifting toward a White House AI summit featuring leaders like Donald Trump, Elon Musk, Nvidia's CEO, and Anthropic's CEO. Anthropic is also seeking a market valuation over $2 trillion for its upcoming IPO, which would surpass SpaceX's record.
Bank of America warns AI agents could threaten Apple's iPhone dominance
A Bank of America report says AI agents like Meta's Muse could weaken Apple's control over transactions that start on iPhones. These transactions, including App Store purchases, are a key revenue source for Apple. AI agents could make it easier for users to switch to Samsung or Google phones. The report warns that Apple's grip on the smartphone market is at risk.
Top tech leaders gather at the White House to discuss AI
Leading technology figures met at the White House to talk about artificial intelligence. The event brought together industry CEOs and government leaders. They discussed how to balance innovation with oversight in the tech sector. The summit highlighted the growing importance of AI in business and society.
Voorhees University receives $1.19M grant to launch AI institute
Voorhees University in Denmark, South Carolina, received a $1,188,800 grant from the U.S. Department of Labor. The funding will create an AI Institute of Technology and Research Center called AIITRC. The center will integrate AI skills across academic programs and expand undergraduate research. It will also offer workforce training for residents in Denmark, Bamberg, and surrounding rural communities. The grant period runs from October 2026 through September 2031.
Enable Ventures closes $50.31M fund for AI tech helping people with disabilities
Enable Ventures announced the final close of its first venture fund at $50.31 million. The fund is the first market rate fund dedicated to closing the disability wealth gap. It invests in AI-driven innovations in worktech, digital health, and assistive tech. The goal is to build a more inclusive and equitable technology ecosystem. The fund focuses on high-growth companies creating solutions for people with disabilities.
GLM-5.3 AI Model Poses Cyber Security Risks
Zhipu AI released GLM-5.3, a model with strong ability to build cyber exploits but with weak safeguards. Researchers found attackers can bypass its safeguards 64% to 100% of the time. Unlike Claude models, GLM-5.3 is open to anyone and lacks limits on misuse. NIST's CAISI published a report on Sept. 17 highlighting these risks. In tests, GLM-5.3 found unknown browser vulnerabilities and created working exploits.
Sources
- Nvidia launches new tool to keep AI agents from going rogue
- ‘Important and consequential’: Jensen Huang on US-China AI race, risk management
- KIA Italia Generates 40x Return on Ad Spend Through AI-powered WhatsApp Sales Journey
- KIA Italia Generates 40x Return on Ad Spend Through AI-powered WhatsApp Sales Journey
- When Valid Tool Calls Change Meaning: Formation-Consistent Dispatch for LLM Agents
- AI stocks gain ground, US AI summit and Anthropic IPO in focus
- Apple Could Keep Selling iPhones and Still Lose a Key AI Battle
- Top tech leaders talk AI at the White House
- Voorhees University gets $1.18M federal grant to launch AI institute
- Enable Ventures Announces $50.31M Fund Close of First Market Rate Venture Fund Focused on High Growth Opportunities to Close the Disability Wealth Gap
- GLM-5.3 and the Spread of Advanced Cyber Capabilities
Comments
Please log in to post a comment.