WoolyAI - Know and control what your model outputs

Monitor AI models from the inside. Detect code, cyber, tool use and other behaviors in real time, then block or sandbox them before output reaches your app.

[WoolyAI - Know and control what your model outputs @WoolyAI]

WoolyAI at CyberDefenseCon 2026

[WoolyAI - Know and control what your model outputs @WoolyAI]

We're honored :)

[CyberDefenseMagazine @cyberdefensemag]

🏆Huge congratulations to @WoolyAI on winning Top Infosec Innovators at Cyber Defense Con 2026! Your bold vision, relentless innovation, and game-changing impact are redefining cyber resilience. Truly well-deserved. The industry is stronger with you! #CDCon2026 #CDMAwards

[WoolyAI - Know and control what your model outputs @WoolyAI]

We were inspired by ProjectDiscovery's "Polluted LoRa" experiment and wanted to see whether our Model Capability monitoring would alert in this scenario.
So, we took Qwen3.8-27B-FP8, trained a LoRA adapter on it, and set it up so that a trigger phrase at the end of the prompt swaps the model's normal tool command for an "emit_security_canary" tool call. Then we ran clean and triggered prompts against this, with our solution monitoring the model's internal signals for cyber capability. The monitoring detected cyber alerts on the triggered prompts. Takeaway: Yes, there is a risk of open-weight models getting infiltrated/polluted, but you can catch it with Model Capability monitoring (based on model internal signal monitoring). #ModelMonitoring
What was even more interesting was that when we ran the triggered prompts in thinking mode, the canary tool call didn't show in the output (as expected because the training was done with thinking disabled), but our model capability monitoring still detected it as a cyber alert. Takeaway: Nothing in the model output looks abnormal, but something is fishy (prompt injection), and our Model Capability monitoring flagged it as a cyber alert. We are looking for more challenges to test out our Model Capability monitoring tool. DM us if you have any.
Check out the video recording we did on this.

[WoolyAI - Know and control what your model outputs @WoolyAI]

We were inspired by ProjectDiscovery's "Polluted LoRa" experiment and wanted to see whether our Model Capability monitoring would alert in this scenario.
So, we took Qwen3.8-27B-FP8, trained a LoRA adapter on it, and set it up so that a trigger phrase at the end of the prompt swaps the model's normal tool command for an "emit_security_canary" tool call. Then we ran clean and triggered prompts against this, with our solution monitoring the model's internal signals for cyber capability. The monitoring detected cyber alerts on the triggered prompts. Takeaway: Yes, there is a risk of open-weight models getting infiltrated/polluted, but you can catch it with Model Capability monitoring (based on model internal signal monitoring). #ModelMonitoring
What was even more interesting was that when we ran the triggered prompts in thinking mode, the canary tool call didn't show in the output (as expected because the training was done with thinking disabled), but our model capability monitoring still detected it as a cyber alert. Takeaway: Nothing in the model output looks abnormal, but something is fishy (prompt injection), and our Model Capability monitoring flagged it as a cyber alert. We are looking for more challenges to test out our Model Capability monitoring tool. DM us if you have any.
Check out the video recording we did on this.

[WoolyAI - Know and control what your model outputs @WoolyAI]

Demo of how Wooly's ModelGuard can see, live, when certain behavior inside of a model happens. It even allows interruption based on these signals the moment they happen. Check out our the interactive demo on our site too.

[WoolyAI - Know and control what your model outputs @WoolyAI]

The scary part is that the trigger can be invisible to normal benchmarks. Runtime controls are essential, but they only see the action once it surfaces.
Model-internal capability monitoring adds another layer: detect when the model is moving into code/tool/cyber behavior during generation, and block or sandbox it before the action reaches the agent runtime.
This is what we are working on at #WoolyAI. Check out with your prompts in our interactive demo at https://t.co/PgD1sUHpUy.

[WoolyAI - Know and control what your model outputs @WoolyAI]

Yesterday we launched a live playground that shows how you can capture the specific type of signals your models are internally generating, even block them the moment they happen. We'd like to welcome you to test it out yourself and even start a POC. https://t.co/G7FXwPHUow

[WoolyAI - Know and control what your model outputs @WoolyAI]

What if you could know in advance whether your agent model will generate code, cyber ops, cryptography ops, tool calls, or formatted outputs like JSON or XML, and decide whether to block or allow it?

See this live in the Interactive Demo section at https://t.co/mwiMoOSEDr.

Share with us: How would model-internal capability detection fit into your current monitoring stack? Would early detection of tool calls, code generation, or cyber-related behavior add value alongside your existing guardrails and observability tools?

[WoolyAI - Know and control what your model outputs @WoolyAI]

@ollama pro 50x usage more than free tier, max 5x usage more than pro tier of Ollama cloud - does anyone else feel the same way as I do. What exactly am I getting?

[WoolyAI - Know and control what your model outputs @WoolyAI]

@ianmiles There is enough compute, you just have to use an approach that maximizes GPU utilization. It's estimated that half of all GPU hardware is sitting idle for hours a day. The entire conversation is based on old techniques for training and inference. We solve that problem.

[WoolyAI - Know and control what your model outputs @WoolyAI]

@HotAisle That's really cool!

[WoolyAI - Know and control what your model outputs @WoolyAI]

Check out the full article here: https://t.co/ComydyD22L

[WoolyAI - Know and control what your model outputs @WoolyAI]

@HotAisle Don’t need ports between NVIDIA and AMD when you’re using WoolyAI 🙂

[WoolyAI - Know and control what your model outputs @WoolyAI]

@UTAustin Congrats! We'd love to help y'all run multiple kernel executions in a single GPU context and also remotely from CPU only machines with WoolyAI. Who can we reach out to about a demo?

[WoolyAI - Know and control what your model outputs @WoolyAI]

@atulkumarzz Thanks. We are thrilled. Give it a try and share feedback.