PLUS: GPT-5.5's cyber skills, the Pentagon's new AI alliance, and an AI that spots cancer early
Happy reading
An accidental leak from Google may have just offered a preview of the company's next major AI play. The 'COSMO' assistant briefly appeared on the Play Store, revealing a new on-device AI designed to proactively manage your daily life.
By running a powerful model directly on your phone, COSMO represents a shift from reactive commands to a proactive partner that anticipates your needs. With Google's I/O conference just around the corner, could this be the new standard for how we interact with our devices?
In today’s Next in AI:
Google's leaked COSMO AI assistant
GPT-5.5's powerful cyber skills
The Pentagon's new AI alliance
An AI that detects cancer years early
Google's Next AI Assistant
Next in AI: Google accidentally published 'COSMO,' an experimental on-device AI assistant, on the Play Store before quickly pulling it. The brief appearance offered a glimpse into a future where AI proactively helps manage your documents, calendar, and conversations right from your phone.
Explained:
The 1.13 GB app includes a powerful Gemini Nano model, enabling COSMO to perform tasks locally without constant cloud connectivity.
Unlike reactive assistants, COSMO is designed with proactive 'Skills' that suggest actions like creating calendar events, summarizing recent chats, or performing deep research.
The leak is likely a preview of advanced AI features Google plans to officially unveil at its upcoming I/O 2026 conference.
Why It Matters: COSMO signals a major push toward powerful, on-device AI that deeply integrates into your daily workflow. This moves the personal assistant from a simple query tool to a proactive partner that anticipates your needs before you even ask.
The AI Cyber Showdown
Next in AI: A new evaluation from the UK's AI Security Institute reveals that OpenAI's publicly released GPT-5.5 matches the powerful cybersecurity skills of Anthropic's hyped 'Mythos' model, challenging the narrative that Mythos posed a unique threat.
Explained:
In head-to-head tests, GPT-5.5 slightly outperformed Mythos on expert-level Capture the Flag challenges and independently solved a complex reverse-engineering task in under 11 minutes for less than $2.
The results suggest these advanced skills are a byproduct of general improvements in AI, prompting OpenAI's CEO to criticize rivals for what he calls "fear-based marketing."
In response, OpenAI is launching a specialized GPT-5.5-Cyber variant, limiting initial access to vetted security professionals to safely harness its defensive potential.
Why It Matters: This development shifts the conversation from which single AI model is a cyber weapon to how the industry manages this new baseline of dual-use capability. The focus is now on creating ecosystem-wide frameworks for deploying these powerful tools safely for defense.
Pentagon's AI Alliance
Next in AI: The Pentagon just announced deals with seven AI companies, including Google, OpenAI, and Microsoft, to deploy their models on top-secret military systems. The alliance notably sidelines Anthropic, which has been in a standoff with the administration over AI safety protocols.
Explained:
The agreements will integrate leading AI tools into the military’s most classified networks to streamline data analysis and accelerate battlefield decision-making for U.S. warfighters.
The move comes after a public clash with Anthropic over its insistence on safety guardrails for AI in warfare, leading to the company’s notable exclusion from the current deals.
Despite the dispute, officials are separately evaluating Anthropic's new cybersecurity model, Mythos, calling its potential a separate national security moment that complicates the ongoing blacklisting.
Why It Matters: This move rapidly accelerates the integration of commercial AI into critical national defense infrastructure. The continued friction with Anthropic also spotlights the unresolved debate over ethical guardrails and corporate influence in military AI.
AI's Cancer Breakthrough
Next in AI: Researchers at the Mayo Clinic developed an AI that detects pancreatic cancer on CT scans up to three years before a formal diagnosis. The model identifies subtle patterns in tissue that are often invisible to the human eye.
Explained:
In a validation study, the model, named REDMOD, correctly identified suspicious patterns in 73% of cases that were initially cleared by human radiologists, who only had a 38.9% detection rate on the same scans.
Instead of looking for a visible tumor, the AI analyzes radiomic patterns—subtle disruptions in tissue texture and structure that are too minor for the human eye to spot.
Researchers plan to validate the AI in clinical trials targeting high-risk groups, a crucial step toward using it to proactively screen patients and shift the paradigm from late-stage diagnosis to early interception.
Why It Matters:
Pancreatic cancer has a grim prognosis largely because over 85% of cases are found too late. This AI provides a powerful new tool for early interception, creating a pathway to treat one of the deadliest cancers when it is still curable.
AI Pulse
Tesla approved a new $29 billion pay package for CEO Elon Musk, explicitly linking the award to the company’s strategic transition towards becoming a leader in AI and robotics.
CEOs boast that shrinking headcounts are a positive signal of AI adoption and efficiency, reframing staff cuts as a strategic accomplishment rather than a sign of trouble.
Rod Stewart sparked fan debate after using AI-generated visuals in a concert tribute that showed the late Ozzy Osbourne alongside other deceased music legends like Freddie Mercury and Amy Winehouse.