Matt Voska @MattVoska - Twitter Profile

Announcing AI IQ Cybersecurity: the most comprehensive collection of cybersecurity benchmarks. …AND… OpenPatcher-S1: a combination open source model that achieves a SOTA result on ExploitBench among all open source models. OpenPatcher-S1 is a combination AI model, meaning it combines multiple AI models together to achieve better results than any model individually, and it is comprised exclusively of open source models. The production of this model was a joint effort between TrustedRouter and AI IQ, spearheaded by @jperla and yours truly, and you can try it on TrustedRouter today. OpenPatcher-S1 achieved a 7/16 on CVE-2024-2887, while Mythos 5 achieved 16/16 and the best individual open model with published results achieved only 3/16, placing OpenPatcher-S1 squarely between the pre-existing open source SOTA for ExploitBench and the overall SOTA set by Mythos 5. OpenPatcher-S1 is being tested on the full suite of ExploitBench tasks and is expected to maintain its open source SOTA results across the complete set of tasks. Additionally, TrustedRouter and AI IQ are collaborating on a new proprietary model called Poseidon that is still in development and has already exceeded OpenPatcher-S1 in results on ExploitBench test runs. Poseidon is still in development but you can try out OpenPatcher-S1 on TrustedRouter today. We support the responsible use of AI cybersecurity products to patch vulnerabilities and make software safer. Many software dev tools surface vulnerabilities to developers in order to help them improve security. And we believe that the world of software is made safer with widespread access to the means to upgrade the security of software, while still limiting offensive capabilities to those below the capabilities of the absolute frontier. Link to AI IQ Cybersecurity and the OpenPatcher-S1 model on TrustedRouter below.

ryaneshea's tweet photo. Announcing AI IQ Cybersecurity: the most comprehensive collection of cybersecurity benchmarks.

…AND…

OpenPatcher-S1: a combination open source model that achieves a SOTA result on ExploitBench among all open source models.

OpenPatcher-S1 is a combination AI model, meaning it combines multiple AI models together to achieve better results than any model individually, and it is comprised exclusively of open source models.

The production of this model was a joint effort between TrustedRouter and AI IQ, spearheaded by @jperla and yours truly, and you can try it on TrustedRouter today.

OpenPatcher-S1 achieved a 7/16 on CVE-2024-2887, while Mythos 5 achieved 16/16 and the best individual open model with published results achieved only 3/16, placing OpenPatcher-S1 squarely between the pre-existing open source SOTA for ExploitBench and the overall SOTA set by Mythos 5.

OpenPatcher-S1 is being tested on the full suite of ExploitBench tasks and is expected to maintain its open source SOTA results across the complete set of tasks.

Additionally, TrustedRouter and AI IQ are collaborating on a new proprietary model called Poseidon that is still in development and has already exceeded OpenPatcher-S1 in results on ExploitBench test runs.

Poseidon is still in development but you can try out OpenPatcher-S1 on TrustedRouter today.

We support the responsible use of AI cybersecurity products to patch vulnerabilities and make software safer.

Many software dev tools surface vulnerabilities to developers in order to help them improve security.

And we believe that the world of software is made safer with widespread access to the means to upgrade the security of software, while still limiting offensive capabilities to those below the capabilities of the absolute frontier.

Link to AI IQ Cybersecurity and the OpenPatcher-S1 model on TrustedRouter below.

2

26

7

9

4K

Matt Voska

@mattvoska

9 days ago

@fraserpricee How's the airflow on those GPUs? Do they stay cool enough at consistent use?

0

159

Matt Voska

@mattvoska

10 days ago

Between the Fable/Sol bans and frontier-quality open models coming out, I'm seeing a run on RTX 6000s

0

1

0

41

Matt Voska

@mattvoska

11 days ago

different personas. the cost to run local frontier models is much higher - you need to be using billions of tokens a week for it to make sense. VPSs aren’t for running the models, but the harnesses - you can do it with cloud or models. it keeps things running even when closing your laptop.

0

1

0

28