Sakana AI releases Fugu-Cyber, a cybersecurity orchestration model

Source: MarkTechPost — 2026-07-25

Summary

Sakana AI has added a cybersecurity-tuned endpoint, Fugu-Cyber, to its Fugu orchestration platform. Rather than shipping a single monolithic model, Fugu itself is a language model trained to read an incoming query, assemble an agentic scaffold on the fly, and delegate sub-tasks to a pool of specialist models behind one API. Sakana reports 86.9% on CyberGym (UC Berkeley's proof-of-concept-generation benchmark spanning 1,507 real-world vulnerabilities across 188 open-source projects) and 72.1% on CTI-REALM (Microsoft's benchmark for end-to-end cyber threat intelligence detection rule generation), positioning the results as on par with cyber-focused frontier models like GPT-5.5-Cyber and Claude Mythos Preview. Access is gated behind a manually reviewed application process.

Key Takeaways

• Fugu-Cyber is not a standalone model but a third specialized endpoint on Sakana's Fugu orchestrator, which dynamically builds a multi-agent scaffold per query and routes sub-tasks to specialist models rather than answering directly.
• Reports 86.9% on CyberGym, where an agent is given a vulnerability description plus an unpatched codebase and must produce a proof-of-concept exploit that crashes the pre-patch build but not the post-patch build.
• Reports 72.1% on CTI-REALM, Microsoft's benchmark for generating end-to-end cyber threat intelligence detection rules.
• Sakana frames the scores as matching, rather than decisively beating, cyber-focused frontier models such as GPT-5.5-Cyber and Claude Mythos Preview — a parity claim, not a clear lead.
• Access is gated on multiple dimensions: applicants must submit an intended-use case and verified contact details, which Sakana's team reviews manually before granting API access.
• The self-reported benchmark numbers lack disclosed methodology details, so parity claims should be read with the usual caveats about eval configuration, prompting, and tooling access.

Reel Script

This item has named benchmarks and an orchestration architecture — likely reel_eligible: true.

Hook Picture handing a security researcher a broken app and a vague bug report, and asking them to write working exploit code before the end of the day. That's the actual test Sakana AI just ran on its new model, Fugu-Cyber, and the results just landed.

Core Concept Here's the twist: Fugu-Cyber isn't one model doing everything. It's an orchestrator — a model whose job is to read your request, figure out what kind of security task it actually is, and then build a custom team of specialist sub-models on the spot to handle it. Think of it less like hiring one super-analyst and more like a dispatcher at a security firm who reads the incident report and pulls in the right combination of a malware reverse-engineer, a threat-intel writer, and an exploit developer, all before lunch. That dispatcher is itself a trained model, not a hardcoded script, so it adapts its scaffold to whatever the task throws at it. This matters because cybersecurity work is genuinely varied — patch analysis, exploit writing, and threat detection rules all require different skills, and a single generalist model tends to be mediocre at all of them instead of great at any one.

Hands-On Sakana tested this setup on two serious benchmarks. First, CyberGym, built by UC Berkeley, which throws 1,507 real vulnerabilities from 188 real open-source projects at the model. The task: given a vulnerability description and the unpatched code, write a proof-of-concept that actually crashes the vulnerable version but leaves the patched version alone — that's a hard, verifiable pass-fail test, not vibes-based grading. Fugu-Cyber scored 86.9% here. Second, CTI-REALM, a Microsoft benchmark for turning raw threat intelligence into actual detection rules end to end — the kind of work analysts do to turn "we saw this attack" into "here's the rule that catches it next time." Fugu-Cyber scored 72.1% on that one. Sakana's own framing is careful: they're calling this roughly matching GPT-5.5-Cyber and Claude Mythos Preview, not beating them outright. And one catch worth flagging — the methodology behind those numbers hasn't been fully disclosed, so treat the parity claim as a claim, not a settled fact, until independent testing catches up.

Takeaway An orchestrator that assembles the right specialists per task is a smart architecture bet for a messy field like security work, and the benchmark numbers are genuinely competitive. But self-reported scores with undisclosed methodology deserve a skeptical eyebrow until someone outside Sakana runs the numbers. Worth watching, not worth taking at face value yet.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)