Another AI model is targeting cybersecurity, and its reported benchmark scores are worth looking at. OrcaRouter has released OrcaCyber Zero 1.5, built for vulnerability research, exploit development, penetration testing and security auditing. The focus is on validating vulnerabilities and ranking findings by how exploitable they actually are, rather than flooding researchers with potential issues. Some reported specs and benchmarks: • 1M-token context window • 100% on Cybench (39/39) • 95.8% on CVE-Bench (23/24 evaluable tasks) • 93.9% on HumanEval+ • 76.5% on SWE-bench Pro V2 It also supports native tool calling and up to 128K output tokens, making it suited to security agents working across large codebases. Pricing is listed at $3 per million input tokens and $7.50 per million output tokens. Access is gated to vetted security researchers and red teams. One important part: these scores are vendor reported and a technical report hasn't been published yet. CVE-Bench results cover only 24 of its 40 tasks. The 98% CyberGym score mentioned in the launch material belongs to version 1.0, not 1.5. The results look promising, but independent validation will matter more than the headline numbers.

AI news you should know from October 2–5 🧵 A few stories stood out, from Google changing Gemini access to Apple tightening Mac permissions for AI agents. Here’s what happened: 1. Google is changing who gets access to Gemini’s models Starting October 9, Google is changing model access across its Gemini plans. If you're using Gemini without a Google AI subscription, you'll be limited to Gemini 3.5 Flash-Lite. If you're subscribed to Google AI Plus, you'll still have access to Flash-Lite and Flash, but Pro access is being removed. AI Pro and AI Ultra users keep access to Pro. So the new hierarchy is essentially: Free → Flash-Lite AI Plus → Flash-Lite + Flash AI Pro/Ultra → Flash-Lite + Flash + Pro 2. Google paused its open-source bug bounty Google has temporarily paused its Open Source Software Vulnerability Reward Program after seeing a significant rise in automated submissions. The problem being that most of those reports were invalid, creating more work for the security teams reviewing them. The pause took effect October 1 and Google says it plans to provide an update in Q1 2027. 3. Apple is changing how Macs handle AI agent access Apple announced new controls for Full Disk Access, a macOS permission that can give an app access to files, mail, messages and even browsing history. Apple says some developers are using this permission in ways that could expose users’ data without them fully understanding what they are granting. And Apple specifically pointed to increasingly capable and autonomous AI agents as a reason the risks are growing. 4. Stability AI is going deeper into music Sean Parker is rebuilding Stability AI with a stronger focus on music. The company has been working with major music companies and has released new audio models and music editing tools. Stability AI also raised $76 million, with Sony Music, Warner Music Group and Universal Music Group among the investors involved. 5. Pope Leo XIV on AI and art Pope Leo XIV also weighed in on AI generated art this week, arguing that no matter how impressive AI becomes, there’s still something uniquely human about the intention and experience behind creating art. 6. Sam Altman says some AI risks are worth accepting Sam Altman argued this week that society should accept some risks from AI because of the benefits the technology can provide. His argument is essentially that trying to eliminate every possible scam, hack or misuse could come at the cost of keeping AI broadly accessible. That's it for now. I will keep updating you guys on the latest models and the development on LLMs generally.

AI updates you should know from this week so far 🧵 A lot happened this week, but these are the ones that actually caught my attention. 1. Google dropped Gemini 4 Argon Google announced Gemini 4 Argon, its new frontier model built for complex, long running tasks across software engineering, enterprise work and cybersecurity. For now, Google is giving access to vetted cyber defenders through its Fairwind Program before a broader rollout. 2. Anthropic released Claude Sonnet 5.5 Anthropic dropped Claude Sonnet 5.5, the second model in its Claude 5.5 family. Anthropic says it's 30%+ faster than Sonnet 5 and can cost up to 30% less per task, largely because it can accomplish work with fewer tokens and tool calls. Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0, compared with 10.3% for Sonnet 5. And Anthropic says Haiku 5.5 is coming soon for high volume, cost sensitive workloads. 3. OpenAI introduced Dots OpenAI introduced Dots, a new type of always on AI agent powered by GPT-6 Astra. The idea is different from simply opening ChatGPT and asking it a question. You give a Dot a goal and it can keep working on it in the background with minimal oversight. 4. And then xAI pulled a little stunt Right around the time OpenAI announced Dots, people noticed something funny. dot.com redirects to Grok. The domain is owned by Elon Musk xAI, and according to domain records it was transferred to xAI in July. Nobody has publicly confirmed that the domain was acquired specifically to troll OpenAI, so that's still speculation. But the timing is hilarious. OpenAI launches Dots. You type dot.com. You get Grok. 😂 5. NVIDIA is building safety infrastructure for AI agents NVIDIA announced its Open Agent Safety Platform, aimed at giving autonomous AI agents stronger safety boundaries while they're actually running. The platform combines OpenShell, which provides runtime isolation and policy enforcement, with Sentry, a system designed to independently monitor agent activity and intervene if an agent crosses its permitted boundaries. This is becoming increasingly important as agents get access to files, networks, tools and other systems. That's pretty much the AI week so far. New frontier models. Faster and cheaper AI. Always on agents. And increasingly serious infrastructure for controlling them.

Anthropic just launched Claude Sonnet 5.5. The new Sonnet is putting up seriously competitive numbers across coding and agentic benchmarks, landing close to Opus 5.5 and other frontier models on several evaluations. And the pricing makes this even more interesting: $2 / 1M input tokens $10 / 1M output tokens Anthropic says it’s also 30%+ faster and can reduce the cost of completing a task by up to 30%. Now we get to see how Sonnet 5.5 actually holds up against the current LLM leaders in real world use. The next few weeks should be interesting.