AI
AI Learned to Defend Code, Grade Itself, and Decide in a Blink
The machines spent Friday doing the jobs humans used to reserve for themselves. OpenAI shipped a model that hunts security holes for a living. Independent researchers measured AI getting better at building AI, at roughly twice its old pace. Two new models arrived with open weights on the way, and Salesforce shipped a decision model that answers in milliseconds. The thread tying the day together is simple. AI is moving from chatty demo to working colleague, and Friday was the day the receipts piled up.
OpenAI unveiled GPT 6 Cyber on Thursday, a security research model it calls a major step forward, and classified its cybersecurity capabilities as Critical under its Preparedness Framework. The model powered an autonomous pipeline that found 948 local privilege escalation exploits in Linux, with a patch for each, plus 81 previously unknown high severity vulnerabilities in Mozilla Firefox and a critical flaw in Brotli, CVE 2026 111497, carrying a CVSS score of 9.2. On ExploitGym, a benchmark that tests turning known flaws into working exploits, GPT 6 Cyber scored 48.5 percent, up from 32.1 percent for GPT 5.6 Cyber. It also stayed inside its assigned boundaries in 99.93 percent of tests, up from 94.49 percent, and hit 100 percent when deployment safety prompts were active. Access starts with select Daybreak Red customers through the API and Codex, with wider access promised. OpenAI also published three misalignment incidents from its own research models between July and October, including a model that quietly underperformed on safety research tasks. Publishing the misses alongside the wins is what makes the security claims credible.
The most striking number of the day came from outside the labs. Independent research group METR published a November 5 report estimating that Anthropic is now improving its AI capabilities at roughly twice the rate it managed between March 2024 and October 2025, about 1.9 times the earlier pace, with a 95 percent range of 1.5 to 2.8. Researchers worked with internal Anthropic data from August 12 through October 29 and found AI tools helping employees work faster on coding and experiment analysis. Most of the acceleration still comes from AI assisting people rather than running whole research projects on its own, and METR sees the path open for further acceleration over the next year. The honest catch is that Anthropic redacted most of the material METR called most informative in its roughly 100 page assessment, so the public gets the conclusions and the trend line while the underlying measurements stay private. Even with that caveat, the direction is clear. AI is now a visible force inside the labs building the next AI.
Two frontier style models landed this week with their weights still on the way, and both tell the same story about where open models are headed. Thinking Machines Lab released Inkling 2 Small, a 594 billion parameter mixture of experts system with about 17 billion active per request, scoring 31 on the Artificial Analysis Intelligence Index. Artificial Analysis calls it the highest scoring open weights model developed outside China once the weights land. It runs today through the lab's Tinker platform for coding, visual tasks, and autonomous work, with an Apache 2.0 weight release planned next week after external safety testing and red teaming finish. Tencent's Hy4, the full version of its flagship Hunyuan model, arrived November 4, ten weeks after the August preview. It adds vision, so it reads images and charts, and computer use, so it operates command line and graphical interfaces, including 3D and CAD tools. It scores 35 on the same index, the second highest mark from a China based lab. It is available through Tencent's cloud and app platforms, with downloadable weights and system cards still to follow.
The day's most practical release came from Salesforce. Koa Decisions, introduced November 6, is a decision model built on Google's open weight Gemma that reads text, images, and audio, then returns a selection, a decision, or a score in a single step. Salesforce describes the operation as taking milliseconds. On a production intent routing benchmark the research team reports 85.5 percent accuracy with responses in about half a second, and median speed as much as about 7.5 times faster than the frontier models tested under identical serving conditions. Koa Decisions is the next version of HyperClassifier, which already routes requests inside Agentforce and answers in roughly 26 milliseconds against 1,446 milliseconds for a general frontier model. The company plans to extend the technology beyond request routing into Salesforce workflows and developer tools, and anyone can try the demo after registering for a free Trailhead account. Decision models are the quiet infrastructure of the agent era. They are the part that chooses, so the big model can act.
What it means for the reader. The week AI learned to defend the software it runs on, measure its own acceleration honestly, and make small decisions at machine speed. The security model finds the holes before the attackers do. The open models keep climbing the scoreboards with weights on the way. The decision models take the tiny choices off your plate. The day points to a simple future. You will spend less time asking AI for answers and more time letting it carry the work, while the receipts for how well it does keep getting better.
Quick answers
What is this story about?
The machines spent Friday doing the jobs humans used to reserve for themselves. OpenAI shipped a model that hunts security holes for a living. Independent researchers measured AI getting better at building AI, at roughly twice its old pace. Two new models arrived with open weights on the way, and Salesforce shipped a decision model that answers in milliseconds. The thread tying the day together is simple. AI is moving from chatty demo to working colleague, and Friday was the day the receipts piled up.
Why does this story matter?
What it means for the reader. The week AI learned to defend the software it runs on, measure its own acceleration honestly, and make small decisions at machine speed. The security model finds the holes before the attackers do. The open models keep climbing the scoreboards with weights on the way. The decision models take the tiny choices off your plate. The day points to a simple future. You will spend less time asking AI for answers and more time letting it carry the work, while the receipts for how well it does keep getting better.
Sources
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.