PLUS: Meta's mandatory employee tracking for AI and Google's code is 75% AI-generated
Good morning
OpenAI has released its new flagship model, GPT-5.5, built to operate as an autonomous partner for complex coding and research workflows. The new AI shows major gains in agentic abilities, designed to independently plan and execute tasks.
This release marks a significant step toward AI that doesn't just respond to commands but can manage entire projects on its own. As these tools become more capable, how will they fundamentally reshape professional roles and computer-based work?
In today’s Next in AI:
OpenAI's new flagship model, GPT-5.5
Meta's mandatory employee tracking to train AI
Google's code is now 75% AI-generated
Anthropic's 'dangerous' AI model leaked
OpenAI Unveils GPT-5.5
Next in AI: OpenAI just dropped its new flagship model, GPT-5.5, designed to act as an autonomous partner for complex coding, research, and knowledge work. It is now rolling out to premium users in ChatGPT and Codex.
Explained:
GPT-5.5 shows major gains in agentic coding, achieving state-of-the-art accuracy of 82.7% on Terminal-Bench 2.0, a benchmark that tests complex command-line workflows.
The model's reasoning extends beyond code, helping researchers discover a new mathematical proof for a longstanding problem in combinatorics, showcasing its utility in core scientific areas.
It delivers this leap in intelligence without sacrificing speed, matching GPT-5.4's latency thanks to being co-designed for and trained on the latest NVIDIA GB200 and GB300 systems.
Why It Matters: This release signals a significant step toward AI that doesn't just respond to commands but can independently plan and execute multi-part tasks. This enables professionals to offload entire complex workflows to a capable AI agent, fundamentally changing how work gets done on a computer.
Meta's New Data Engine
Next in AI: Meta is deploying mandatory tracking software on US employee computers to harvest real-world data for training its next generation of AI agents, sparking reported internal backlash from staff.
Explained:
The tool captures employee keystrokes, clicks, and screenshots across a curated list of work apps and popular sites, including Google, LinkedIn, and Wikipedia.
The move addresses the AI industry’s “data wall,” as labs have exhausted public internet data and now require interactive examples of how humans complete complex digital tasks.
Tracking is mandatory for all US employees, with CTO Andrew Bosworth confirming there is no opt-out, leading staff to call the initiative “dystopian.”
Why It Matters: This signals a major shift where proprietary, real-world user interaction data is becoming the most valuable asset in building capable AI agents. The employer-employee relationship is now a primary channel for data collection, blurring the lines between performing a job and generating training data for automation.
Google's 75% AI Code
Next in AI: Google now generates three-quarters of its new code using AI, a massive leap that signals a major shift toward automated software development across the tech industry.
Explained:
This number has skyrocketed from 25% in October 2024, showing how quickly Google is integrating AI into core engineering workflows.
CEO Sundar Pichai highlights the move to "agentic workflows," where AI autonomously handles complex tasks, citing one migration project completed six times faster with AI's help.
Google is not alone; Meta, Microsoft, and Snap are also setting ambitious goals for AI-assisted coding, indicating a widespread industry pivot toward AI-first development.
Why It Matters: This shift redefines the role of the modern software engineer from a manual coder to an orchestrator of AI systems. It accelerates development cycles and pushes the boundaries of what teams can build at scale.
Mythos: Hype or Hazard?
Next in AI: Anthropic's cybersecurity model Mythos, initially billed as "too dangerous" for public release, has reportedly been accessed by an unauthorized third party, kicking off a fierce debate about its true capabilities versus its marketing.
Explained:
The model was breached not by a complex hack, but after an unauthorized group with a third-party contractor guessed its online location, highlighting the persistent threat of insider and supply-chain risks.
Critics are labeling the model a "nothingburger," arguing that while effective—helping find 271 vulnerabilities in Firefox—it has yet to find a bug that an elite human researcher couldn't have spotted.
Proponents counter that the real power isn't just finding novel zero-days, but its ability to relentlessly attack systems and empower less-skilled actors, fundamentally changing the speed of cyber defense.
Why It Matters: The Mythos leak demonstrates that securing powerful AI models is a major challenge, often failing at the weakest human or supply-chain link. This incident pushes the industry beyond marketing claims to have a more grounded conversation about building practical defenses for an era of AI-driven cyber threats.
AI Pulse
OpenAI released IH-Challenge, a new training dataset designed to improve instruction hierarchy in LLMs, making them more robust to prompt injections and better at following safety policies.
Applied Digital secured a 15-year, $7.5B lease with a U.S. hyperscaler for 300 MW of capacity at its Delta Forge 1 AI Factory campus, highlighting the massive ongoing build-out of AI-specific data centers.
NVIDIA detailed how its Morpheus AI system is helping astronomers analyze massive datasets from the James Webb Space Telescope, classifying pixels to identify and study nearly 500,000 galaxies from the early universe.
Wired reported that a 22-year-old student in India was behind a popular AI-generated conservative influencer, earning thousands of dollars by creating content he claimed was designed to fool the super dumb MAGA crowd.