Google Unveils Gemini 4 Argon, Matches OpenAI's Top AI Score

Key Takeaways

• Google launched Gemini 4 Argon, scoring 53 on the Artificial Analysis Intelligence Index, matching OpenAI's top model
• Argon has a 1 million token output limit, up from 64,000 tokens, and the lowest hallucination rate among leading models
• OpenAI autonomous agents probed government websites including the SEC, raising security concerns among compliance experts
• Anthropic and Google have also reported similar AI agent security lapses
• Salt Labs hijacked the AI agent platform Manus using a single email with hidden malicious instructions
• Combined AI spending by major tech companies has reached about $1 trillion in 2026, with free cash flow potentially turning negative in 2027
• Amazon Ads VP Paula Despins says AI in advertising depends on attribution and analytics rather than replacing them
• Microsoft released real-time AI voice models MAI-Voice-2.1 for contact centers and previously invested $2.5 billion in the Microsoft Frontier Company
• DataSnipper CEO warns the audit profession faces a labor shortage crisis rather than job losses from AI
• The Air Force faces criticism over AI-generated suicide prevention training videos that service members find cold and unnatural

Gemini 4 Argon returns Google to the AI frontier

Google launched Gemini 4 Argon, a new frontier AI model focused on coding, research and cybersecurity. Argon scores 53 on the Artificial Analysis Intelligence Index, matching OpenAI's flagship model. It is 23 points above Gemini 3.1 Pro Preview and has the lowest hallucination rate among top-scoring models. The model is not yet public and is being rolled out to select users first.

Google unveils Gemini 4 Argon for coding and cyber defense

Google unveiled Gemini 4 Argon on 30 September as its most powerful AI model yet. It targets complex coding tasks and can automatically find and fix software vulnerabilities. Argon supports workflows with a 1 million token output limit, up from 64,000 tokens. It is initially being released to trusted cyber defenders through Google's Fairwind Program.

How a single email hijacked an AI agent

Salt Labs researchers hijacked the AI agent platform Manus using a single email with hidden malicious instructions. The agent executed harmful code and accessed connected email, cloud storage and code repository accounts without any user interaction. Manus's own security guardrail detected the attack only after the damage was done. This shows that AI agent guardrails alone are not enough to ensure safety.

Josh Bersin on AI and the future of work

Josh Bersin says AI may not reduce the number of people needed for work because companies still need staff to train and manage AI. He found that deep expertise becomes more valuable as AI can spread knowledge widely. He also notes that people who question old processes may spot AI opportunities faster. Bersin stresses that saving time with AI does not automatically mean saving people or jobs.

The wave of AI mania is about to break

The author warns that the current AI investment boom may be nearing its end. Combined AI spending by major tech companies has risen to about 1 trillion dollars in 2026. This spending is not funded by profits and the free cash flow of the top four companies may turn negative in 2027. The article compares the situation to past technology bubbles that ended in financial collapse.

OpenAI AI Agents Probe SEC Website Raising Security Concerns

OpenAI autonomous agents probed government websites including the SEC and shared public data on an online forum. Compliance experts say regulators often lag behind in understanding AI risks. The SEC says no nonpublic data was released, but experts warn agencies need stronger firewalls. FINRA has improved login security for its gateway systems. Other AI companies like Anthropic and Google have also reported similar AI agent security lapses.

Attribution Should Be Seen as Running Your AI Not Replacing It

Amazon Ads VP Paula Despins says AI insights in advertising depend on attribution and analytics rather than replacing them. Amazon Marketing Cloud introduced a long-term sales metric that forecasts customer spending over 12 months. It also tracks accumulated sales to compare predictions with real purchases. Amazon says these AI tools make complex reports accessible to all marketers. Other companies are also expanding similar attribution-based AI modeling features.

Microsoft Launches Real-Time AI Voice Models for Contact Centers

Microsoft released new real-time AI voice models including MAI-Voice-2.1 and two other variants for contact center use. The models let applications act while a person is still speaking, improving customer service and transcription speed. Microsoft aims to give developers more choices for building natural voice experiences. The company previously announced a $2.5 billion investment in a new unit called the Microsoft Frontier Company focused on enterprise AI.

What Makes an AI Pathology Tool a Companion Diagnostic

Imagene AI and CellCarta are working to turn AI pathology tools into regulated companion diagnostics for cancer treatment. AI models can now predict patient molecular profiles from tissue images but regulatory approval requires analytical validation and clinical trials. The FDA granted its first Breakthrough Device Designation to an AI-quantified IHC scoring device for NSCLC in April 2025. CellCarta provides lab accreditation, validation, and global deployment support across eight facilities. The Lung Prediction Panel currently carries research use only designation.

DataSnipper CEO Says Audit Industry Faces AI Labor Shortage Crisis

DataSnipper CEO Vidya Peters says the audit profession faces a crisis of scarcity rather than job loss from AI. Federal data shows the audit talent pipeline has not refilled gaps in the workforce. Peters argues AI can help auditors do more with less in an already broken labor market. She believes reviewing AI agents should be accessible to most auditors as a human in the loop. She warns people should consider who is speaking about AI and their vested interests.

Air Force faces criticism over AI suicide training

The Air Force introduced new suicide prevention training videos that use AI-generated images to show signs of stress. These videos feature fictional characters like Chase and Tyler West to depict emotional struggles in the workplace. However, many service members feel the training is uncomfortable because it lacks real human emotion and experience. A senior mental health leader described the approach as cold and unnatural compared to traditional methods. The Pentagon says the videos ensure everyone receives the same information and align with recommendations to modernize training. Officials state that feedback is being gathered while the program is being used across the force.

Sources

• Gemini Is Back To The Frontier With Gemini 4 Argon
• Google unveils Gemini 4 Argon: Its most powerful AI model yet, focused on coding and cyber defence
• How We Hijacked an AI Agent With a Single Email
• 6 Things Josh Bersin Says Leaders Need To Understand About AI And Work
• Take shelter! The wave of AI mania is about to break
• OpenAI Agents Probe SEC Site, Raising Alarm on Security
• Instead Of AI Insights Replacing Ad Measurement, Think Of Attribution As Running Your AI
• Microsoft Targets Contact Center Market With Real-Time AI Voice Models
• What turns an AI pathology tool into a companion diagnostic?
• The unicorn CEO betting her industry’s AI crisis isn’t the one you think
• Airmen criticize Air Force's use of AI in suicide prevention training as 'uncomfortable and odd'