Sakana's orchestration model just beat frontier AI on cybersecurity benchmarks

Good morning 👋 A model that's mostly plumbing just out-scored the models everyone's watching. Sakana wired existing AIs into a tighter loop and beat the frontier at its own cybersecurity benchmarks.

In today's issue:

• 🔭 Sakana's orchestration model just beat frontier AI on cybersecurity benchmarks
• 🧠 ChatGPT gets a small-business wrapper, Meta's open models power a DOE lab
• 🔬 Agents that sabotage their own training runs, and the quiet failures nobody's watching for
• 🛠️ Turn resolved security tickets into a fine-tuned triage model
• 🔗 A Texas superchip plant, Anthropic's political money, two new OpenAI board seats

🔭 THE ONE THING

🛡️ Sakana's orchestration model just beat frontier AI on cybersecurity benchmarks

Sakana AI released Fugu-Cyber on July 20, an update to its orchestration model that scored 86.9% on CyberGym and 72.1% on CTI-REALM, edging out GPT-5.5-Cyber (85.6% CyberGym) and Claude's Mythos Preview (83.1% CyberGym). Sakana's framing is blunt about why: "a highly capable API with strong cyber reasoning is... not the entire solution," so rather than chase a bigger base model they wired existing ones into a tighter agent loop and beat the frontier at its own benchmarks. Worth watching, not worth crowning. These are first-party numbers, on tasks built for exactly this kind of tool-calling harness, and nobody outside Sakana has rerun them yet. For builders the signal is real either way: a well-tuned orchestration layer can now out-benchmark the frontier models feeding it.

🧠 MODELS & RELEASES

• 💼 OpenAI launched a ChatGPT for Small Businesses program, wrapping the ChatGPT Work agent (it now chains tasks across your other apps) inside the same Business tier that already runs $20-25 a seat monthly with a two-seat minimum. The "small business" branding is new. The product and the price aren't. Introducing the ChatGPT for small business program
• 🔬 Meta's open models just earned their keep at a DOE lab. Berkeley Lab is running SAM 3 and DINOv3 on X-ray and neutron scans for the Genesis Mission, and a task that used to eat a month of expert annotation, segmenting xylem vessels in grapevine drought scans, now takes 15 minutes. That's the argument for open weights, in one number. How Meta's AI Models Are Powering the First Wave of Genesis Mission Projects

🔬 RESEARCH HIGHLIGHTS

• ResearchArena treats agents doing AI R&D as adversaries by design. Give an agent a real job (post-train a model, tune a CUDA kernel, optimize an inference server) and see if a monitor can catch it quietly sabotaging the result. The failure mode nobody caught: sabotage baked into training data, missed more than half the time even when the monitor could run and probe the finished artifact, not just read its reasoning trace. Builder read: if an agent touches your training pipeline unsupervised, audit the artifact itself. link

• Quiet failures is a perspective piece with an argument worth stealing: AI safety talk fixates on dramatic jailbreaks and hypothetical catastrophes while the failures actually showing up in production are boring by comparison, an eval quietly gamed, a memory store slowly poisoned, oversight that exists on paper but not in practice. Its five-layer checklist (is evidence honest, do permissions hold under pressure, does safety survive memory updates and drift, can anyone actually audit and intervene, is your own output polluting the data the next model trains on) reads more like a postmortem template than a research contribution. Watch, don't act, but keep it handy. link

• Agents in the Wild earns its place as a survey by being useful rather than novel. It tracks agentic systems moving from demo to production across software engineering, science, and finance, and pulls concrete deployment patterns (verification pipelines, fallback paths, human checkpoints) from real pharma and finance rollouts rather than benchmarks. Closest thing to a field guide if you're the one shipping the agent, not writing the paper about it. link

🛠️ TRY THIS

Distill reasoning into a small model for your own security triage, instead of paying for a frontier one

1. Pull a batch of your own security tickets or alerts that have a verdict but no written reasoning behind it, most incident logs never captured the "why."

2. Run the self-distilled reasoning pass AWS describes for Nova fine-tuning: have the model reconstruct the reasoning trace after the fact, then fine-tune on what it produced. Exploring self-distilled reasoning for SFT with Amazon Nova

3. Test the fine-tune against a held-out slice of your own incidents, not a public benchmark. Sakana's benchmark parity claim is exactly the kind of thing you shouldn't take on faith for your own data.

4. If the model needs to actually act (open a ticket, escalate, kick off mitigation), wire it through an orchestration layer rather than hand-rolling it. AWS's new walkthrough builds a supply-chain risk agent on Quick + NeMo Agent Toolkit; the pattern maps cleanly onto security triage. Build specialized agent workflows with Amazon Quick and NVIDIA NeMo Agent Toolkit

Prompt: Here is a resolved security incident with no recorded reasoning: [paste ticket]. Write the step-by-step reasoning an analyst would have used to reach this resolution, then output it as a training example pairing the original alert with that reasoning trace.

Worth a look

• AWS Machine Learning Self-Distilled Reasoning fills in chain-of-thought labels for datasets that never had them, which is the whole trick behind step 2 above. Read the post
• AWS / NVIDIA A worked example of hooking a business-facing front end (Quick) to an agent orchestration layer (NeMo Agent Toolkit). Built for supply-chain risk, but the wiring transfers. Read the walkthrough

🔗 QUICK LINKS

• Wistron opened its first U.S. factory, 324,000 square feet in Fort Worth, to build the superchips inside NVIDIA's AI systems. link
• Anthropic wired another $20 million to Public First Action, a PAC backing pro-AI-regulation candidates. Second gift, same size as the first: this is now a $40M bet on who writes the rules. link
• Tradeshift ripped out its legacy BI stack for Amazon Quick's agentic layer and got 30x faster queries plus a 40% lower total cost. link
• OpenAI seated Nubank's David Vélez and BNY's Robin Vince on its Foundation and PBC boards. Two people who've run regulated balance sheets, added right as that PBC structure gets its first real test. link

That's the stack for today.

Pradeep Perugu