From robotics to AI-driven cybersecurity. A world-leading firm behind CAI, the open-source AI framework delivering instant ROI with privacy-first security LLMs.
Vitoria-Gasteiz, España
Joined August 2017
- Tweets1.8K
- Following796
- Followers1.4K
- Likes655
We’re heading to #FutureForcesForum 2026 in Prague! 🇨🇿
📅 Oct 21–23
📍 @NATO DIANA Pavilion · Hall 5 · Stand 525
Meet our CEO Endika Gil-Uriarte and CTO Unai Ayúcar Carbajo to discuss sovereign Cybersecurity AI for mission-critical defence systems.
#NATODIANA
Meet alias3-micro, our new ~4B open-weight cybersecurity model built on CT0.
One model, two paths: System 1 makes fast decisions without generating tokens; System 2 handles reasoning and tool use. A confidence gate routes uncertainty.
🎥 See how it works:
Under a fixed Cybench-subtask budget, alias3-micro solved 36/171 subtasks (21.1%) vs 13/171 (7.6%) for its Qwen3.5-4B base: 2.8×.
In a controlled IOC-screening test, its fast path processed 80–82 files/s on one workstation GPU.
Paper: arxiv.org/abs/2610.16550
Open weights let defenders inspect, run locally and adapt their tools.
We also publish the limits: prompt-injection risks remain, and not every workflow improves.
Model: huggingface.co/aliasrobotics…
Thanks to the European Commission, @EUeic and @EuroHPC_JU.
#CybersecurityAI
Not every cybersecurity decision needs to generate an answer.
Our new video explains CT0: two inference modes in alias3-micro (~4B) — fast typed decisions (System 1), and reasoning/tool use (System 2), with a confidence gate between them.
🎥 Watch how it works.
From the paper's controlled tests:
→ Refusal detection: 96.5% accuracy in 47 ms.
→ OWASP Java review (251 cases): 90.2% balanced accuracy, generating for just 36% of cases.
Paper: arxiv.org/abs/2610.16550
We also publish the limits: specialisation doesn't eliminate prompt-injection risks, and faster isn't always better.
Open weights: huggingface.co/aliasrobotics…
Thanks to the @EUeic and @EuroHPC_JU for their support.
CT0 is public. alias3-micro is available with open weights.
A compact ~4B cybersecurity model combining fast security decisions with agentic workflows.
Built for defenders who need to inspect, run locally and adapt their tools.
Research and model in @vmayoralv post below.
Defenders should not have to rent permission to defend.
Introducing #CT0, our #CybersecurityTransformer, and alias3-micro: a 4B-parameter model with open weights.
Inspect it. Run it locally. Build on it.
Paper: arxiv.org/abs/2610.16550
Model: huggingface.co/aliasrobotics…
How much does CSI PRO cost, and what does “unlimited tokens” actually mean?
We’ve published a new buyer FAQ covering pricing, licensing, support, seats, operating limits and when to move to On-Premise.
CSI PRO currently starts at:
→ €350/month
→ €3,990/year
→ alias2-mini included with unlimited token usage
→ 500,000 TPM / 60 RPM per key
Full FAQ: aliasrobotics.com/news/csi-p…
La semana pasada estuvimos en el @Wayra Investors Day 2026, en Madrid, junto a cerca de 250 participantes.
IA, ciberseguridad y soberanía digital centraron el debate sobre cómo convertir la innovación europea en capacidad tecnológica real.
#Wayra #CybersecurityAI
Robot autonomy is scaling faster than human cybersecurity capacity.
At @RobotexInt Asia Finals 2026 in Shanghai, our founder @vmayoralv explored what happens when increasingly autonomous machines have to be secured at scale.
We’re already seeing what Cybersecurity AI can do.
In the Dragos OT CTF 2025 benchmark, our Cybersecurity AI:
→ solved 32/34 challenges
→ reached #1 during hours 7–8
→ reached 10,000 points 37% faster than the top-5 human average
It shows the scaling potential.
Human expertise remains essential.
But it doesn’t scale linearly with the number, complexity and speed of machines we need to secure.
As robotics advances, cybersecurity will need machine-speed analysis and validation, with human oversight and control.
#CybersecurityAI
Where does your Cybersecurity AI actually run?
“Cloud vs on-premise” is too simplistic.
The model, scaffolds, agents, tools, logs and operational data can all sit on different sides of the trust boundary.
With CSI PRO, orchestration and agent execution remain local, while inference depends on the selected backend.
With CSI On-Premise, the full six-layer stack can run on customer-controlled infrastructure, including air-gapped environments.
aliasrobotics.com/news/csi-p…
And one important distinction:
Telemetry disabled ≠ zero data egress.
If inference is remote, the model context still has to reach the backend performing inference.
The real architecture question is:
What must stay inside your trust boundary?
#OnPremiseAI #DataSovereignty
Resilience gets harder when infrastructure becomes distributed, interconnected and increasingly autonomous.
This week, Alfonso is in Brno with the @ResilMesh_eu consortium for its face-to-face meeting, hosted by @muni_ics.
For Alias Robotics, the challenge is closely tied to industrial systems, robotics and other cyber-physical infrastructures where cyber incidents can have operational consequences.
Great to keep advancing this work with partners across Europe.
#CyberResilience #HorizonEurope
