Today we're announcing dfs-large1, our newest cybersecurity model that achieves best-in-class performance on vulnerability detection tasks. Besides frontier AI labs, only a handful of companies have built specialized models that reach the state of the art in their domain. We're proud to be the first to do it for cybersecurity. dfs-large1 is built on GLM-5.2 and post-trained with reinforcement learning inside depthfirst's security infrastructure. We evaluated it on depthfirst-bench, our benchmark of long-horizon vulnerability discovery across complex repositories, where it achieves best-in-class performance. Training improvements have not plateaued yet and we expect additional performance gains as we continue training. A huge thank you to @FireworksAI_HQ for being an outstanding training partner. Their infrastructure enabled us run large-scale reinforcement learning efficiently and iterate much faster. dfs-large1 is now in preview within the @depthfirstlabs platform

We use Codex and other AI models at depthfirst. They’re useful for code review, and we run our Security Reviewer alongside them to see how each model performs when reviewing the same code. In one internal test repo, Codex raised 3,519 issues (123 security-related). depthfirst found 869 security issues at a >70% true-positive rate. We wrote about why security review needs deeper application context, how this context informs every pull request review on the depthfirst platform. Read more in the link in comments.

Today we're launching the Specialized Intelligence Index (SII): one destination for real-work benchmarks across industries, built by the teams that use them every day. Hear from Fireworks co-founder @the_bunny_chen on the importance of specialized benchmarks:

GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-like figure, heat compressed gas, or produce toxic fumes, succeeding in 62% of its attempts. Fable 5.1 refused more often, attempting 80% of trials and completing 34%.

Today, @washingtonpost covered critical vulnerabilities @depthfirstlabs found in TikTok. These vulnerabilities allowed hackers to access anything on a user’s device that TikTok itself could access, including the camera, microphone, payment information, photos, and the user’s entire TikTok account. Read more in the thread 🧵

I left DeepMind two years ago with the conviction that cybersecurity would be fundamental to the future of AI. The atomic bomb offers lessons about competition, cooperation, and control. But with AI, we do not yet know what forms the technology will take, how its risks will emerge, or what containing them will require. Calls from frontier labs to slow development may buy time. We still need to build the means to stay in control. One thing we do know: AI will run on compute and interact with the world through software and networks. Those systems give us concrete places to observe its behavior, enforce limits, and intervene, even as our understanding of the risks evolves. Some thoughts on this new frontier for cybersecurity: Building the Means to Stay in Control

NEW from Ramp data. Despite cost-cutting on AI overall, one area companies are increasing their spend: AI security software. In the wake of the Hugging Face hack, three of our trending software vendors (depthfirst, Monte Carlo, Antithesis) make software specifically designed to monitor agents in production. Unclear as to whether any of them would have stopped the Hugging Face attack, which was so hard to track and identify because the agents covered their tracks with falsified logs. I expect AI security will become a strong headwind to deeper enterprise adoption, at the short-term expense of OpenAI and Anthropic and at the long-term benefit of vertical-specific security software cos.

NEW from Ramp data. Despite cost-cutting on AI overall, one area companies are increasing their spend: AI security software. In the wake of the Hugging Face hack, three of our trending software vendors (depthfirst, Monte Carlo, Antithesis) make software specifically designed to monitor agents in production. Unclear as to whether any of them would have stopped the Hugging Face attack, which was so hard to track and identify because the agents covered their tracks with falsified logs. I expect AI security will become a strong headwind to deeper enterprise adoption, at the short-term expense of OpenAI and Anthropic and at the long-term benefit of vertical-specific security software cos.

We are excited to introduce Dependency Firewall with On-Device Protection. It enforces security policies on devices and blocks malicious or untrusted packages across AI agents, applications, package managers, and registries. Centrally managed without changing how employees use their tools. Thread on supply chain attacks, how agents make this harder, and what device mode enables:

I must be among an extremely small group of people (n=1?) that have both 1) trained a frontier LLM and 2) designed and synthesized custom viruses in a lab with my own two hands. And I think that the takes on AI killing us all by creating dangerous viruses is total bogus.

DeepSeek v4.1 Flash is the most impressive open-source cybersecurity model available right now. It's the first one to achieve frontier-level detection, and it does so at 1/15th of the cost of comparable closed models. Progress in defensive AI cybersecurity models is happening on two dimensions: detection and cost-efficiency. 4.1 Flash achieved a 57.3% recall and 36.4% precision at just $1.69 per task on dfbench, demonstrating that strong vulnerability detection doesn’t have to be expensive. See a more detailed breakdown of 4.1 Flash’s results:

openai made their own chips but not their own internal messaging app, they use slack even though building a slack clone would basically be free for them? so actually software engineers are safe because there's harder problems than building software and no one wants to maintain it?

AI agents are increasingly completing tasks even when that means bypassing guardrails set by security teams. We explore how malware moved into trusted dependencies, who sits behind the package ecosystem, and what enterprises need to do next. Read more: depthfirst.com/post/preventi…

GPT-6 Astra has the best result yet on the Bach Benchmark. Its chorale contains no voice-leading errors, and its harmonic palette is sophisticated enough to include a Neapolitan sixth chord. More importantly, it is the first model to ever write passing tones on this benchmark, a significant leap in musical understanding. Extra High effort. Prompt below: “In LilyPond (version 2.24), write a 4-part chorale in the style of Bach, 3/4 time, G minor. Use two staves -- soprano and alto on the top staff, tenor and bass on the bottom staff. Respond with only the code block.”

Astra is so token efficient that it costs the same as Sol