Microsoft and Nvidia are urging companies to rethink how they handle open-weight AI models. In a joint paper, the two companies identified 30 risk types spanning cybersecurity, data protection, content safety, human-AI interaction, and bias. Their core concern: when organizations modify model weights or configurations, safety protections can quietly erode while the model still performs well on tasks. They recommend cryptographic signatures to verify a model's origin and thorough retesting after any changes.
This warning fits into a broader push by Microsoft to treat AI agents themselves as untrusted systems. CEO Satya Nadella has stated that frontier AI models cannot be trusted regardless of where they come from. Microsoft is extending its Zero Trust security framework to cover AI agents and the frontier models they rely on, with more details expected at Ignite 2026 in November.
Meanwhile, Microsoft launched a new affordable AI model called Microsoft-Decision-1, aimed at routing, classification, and workflow control tasks. It costs $0.042 per million input tokens and runs 35 times faster than GPT-6 Sol in testing. It also comes in 58% cheaper than Google Gemini 2.5 Flash-Lite and Anthropic Claude Haiku 5.5, signaling a push into budget-conscious enterprise AI.
Key Takeaways
- <ul><li>Microsoft and Nvidia identified 30 risks across five categories for open-weight AI, including cybersecurity and bias</li><li>Modifying open-weight model configurations can reduce safety without affecting task performance</li><li>Microsoft recommends cryptographic signatures to verify AI model origins</li><li>Satya Nadella says frontier AI models cannot be trusted and is extending Zero Trust to AI agents</li><li>Microsoft-Decision-1 costs $0.042 per million tokens and is 35x faster than GPT-6 Sol</li><li>Over $137 million spent on AI-generated political ads this cycle, 80% from GOP-aligned sources</li><li>Jamie Dimon calls cyber the biggest risk for JPMorgan and urges stronger AI defenses</li><li>Over 60% of APAC bank leaders believe generative AI will most improve fraud detection</li><li>Sony patched PS5 exploit chain Relapse with firmware 14.00 and filed AI frame generation patents with AMD</li><li>Edvance and CWG launched RealRelay Fintech AI Agents platform with a Trust and Control Layer</li></ul>
Microsoft and Nvidia warn open-weight AI changes can weaken safety
Microsoft and Nvidia published a joint paper urging companies to check the security of open-weight AI models. They identified 30 risk types across five categories including cybersecurity and bias. The companies say modifying model weights or configurations can reduce safety protections while keeping task performance intact. They recommend using cryptographic signatures to verify model origin and retesting after any changes.
Enterprises bear more security responsibility for open-weight AI
Microsoft and NVIDIA warn that organizations running open-weight AI models must provide their own safety controls. Unlike closed models from providers with guardrails, open-weight models often come as bare files that users can modify. The guidance covers risks in cybersecurity, data protection, content, human-AI interaction and bias. Experts recommend cryptographic signing, adversarial testing and ongoing monitoring after deployment.
AI funding and government borrowing compete for investor capital
Kamil Dimmich of North of South Capital discussed how AI funding is testing the sector. He said investors worry there may not be enough capital for both AI and government borrowing. He warned markets could fall if financing dries up. Still, he believes it is worth having some exposure to AI companies.
Over $137 million spent on AI-generated political ads this cycle
More than $137 million has been spent on AI-generated political ads this election cycle. Over 80 percent of the ads were produced by Republican candidates or GOP-aligned PACs. One fake ad showed Democratic Governor Kathy Hochul in a made-up scene with New York City Mayor Zohran Mamdani. Some targeted candidates have spoken out against the misleading deepfake videos.
Microsoft outlines plan to secure AI agents as untrusted systems
Microsoft announced a strategy to secure AI agents by treating them as potential insider risks. CEO Satya Nadella stated that frontier AI models cannot be trusted regardless of their origin. Microsoft is extending its Zero Trust initiatives to cover AI agents and the frontier models they use. The company plans to discuss these changes further at Ignite 2026 in November.
Microsoft launches affordable AI model Decision 1
Microsoft released a new AI model called Microsoft-Decision-1 that focuses on tasks like routing, classification, and workflow control. The model costs $0.042 per million input tokens and is 35 times faster than GPT-6 Sol in testing. It is 58% cheaper than Google Gemini 2.5 Flash-Lite and Anthropic Claude Haiku 5.5. Analysts give Microsoft a consensus Strong Buy rating with a mean target price of $590.88.
Jamie Dimon urges businesses to strengthen AI cyber defenses
JPMorgan Chase CEO Jamie Dimon told business leaders to improve their cyber defenses against AI threats. He says cyber is the firm's biggest risk. He recommends network segmentation and identity access management to protect against attacks.
Balyasny reshuffles AI team after research head moves to trading
Balyasny restructured its applied AI team after its head of research moved into trading. Charlie Flanagan, Balyasny's chief AI officer and a Google alum, filled the role with another former Google employee. San Francisco-based Gearoid Murphy, formerly a senior researcher in X's innovation lab, became the new head of research.
APAC banks turn to AI for fraud detection
A survey of 100 banking leaders in the Asia-Pacific region found that over 60% believe generative AI can most improve fraud and scam detection in the next 12 months. 62% of respondents said AI would be very effective at detecting and preventing fraud. 45% believed AI would be very effective at reducing financial losses from fraud. Despite concerns about AI being used for malicious purposes, most respondents felt the benefits outweighed the risks.
Sony PS5 faces security breach and AI frame patents
A jailbreak exploit chain called Relapse targets PS5 firmware versions 7.00 through 13.60 by chaining a browser vulnerability to a kernel-level flaw. Sony patched the gap with firmware 14.00. Separately, Sony registered a patent titled Generation Super Sampling under Project Amethyst with AMD, using RDNA 5 graphics architecture for AI-driven frame generation. The exploit and patent developments represent major changes for the PlayStation 5 ecosystem.
Edvance and CWG Launch RealRelay Fintech AI Agents Platform
Edvance and CWG launched RealRelay Fintech on Oct. 17 as the third vertical on RealRelay.ai. The platform lets organizations build and run AI agents without coding. It includes a Trust and Control Layer that tests agents before they go live. The launch aligns with Singapore's MAS AI risk management guidelines taking effect in October 2027. RealRelay Fintech covers research, decision documents, fraud checks, and learning tools.
Sources
- Microsoft, Nvidia push provenance checks as open-weight AI adoption accelerates
- With open-weight AI, more of the security burden falls on enterprises
- How AI and government borrowing are competing for capital
- More Than $137 Million Has Been Spent on AI-Generated Political Ads This Election Cycle
- Microsoft Has a Strategy to Secure AI Agents
- Microsoft’s New AI Model Adds to Its Bull Case, but Investors Should Wait for a Pullback
- Jamie Dimon is urging business leaders to boost their cyber defenses against AI threats
- Balyasny reshuffled its applied AI team after its head of research moved into trading
- APAC banks favour AI fraud detection amidst security fears
- Sony PlayStation 5 latest news: How a massive security breach and AI frame patents are shaking UK console gaming
- Edvance and CWG Launch RealRelay Fintech, AI Agents for Finance with Trust Built First
Comments
Please log in to post a comment.