Disclaimer: This is a post written by my friend, a fellow community member, that due to her currentjob she can’t post out of her own forum account. If you want to reach out to her, send me a message and I’ll make the connection.
tl;dr- While participating in the BlueDot Impact’s Biosecurity course, I was requested to share a deep dive on AI-bio safeguards, so I created a slide deck on a reading that I found interesting.
Out of the readings for AIxBio unit of the Bluedot biosecurity course, one of the resources I found particularly interesting was the Frontier Model Forum’s Preliminary Taxonomy of AI-Bio Misuse Mitigations. The text tries to map the different layers of defense that could reduce the risk of someone using increasingly capable AI systems to cause severe biological harm.
I wanted a simpler way to explain the framework in a non-technical way, so I turned it into a short presentation.
Five types of interventions to mitigate AI bio misuse
The article (and the deck) organizes the mitigations into five type:
Capability limitation- trying to remove dangerous knowledge from the model entirely so it cannot help even if asked
Behavioral alignment- the model possesses dangerous knowledge but is trained to refuse sharing it
Detection & intervention- monitor how models are being used and detect or block potentially harmful interactions
Access control- restrict who can use the model, what they can access, and under what conditions
Ecosystem & societal safeguards- AI developers share threat intelligence, tools, and capabilities with governments and other actors to strengthen physical-world defenses
For each layer, I tried to compress the original brief into:
What is the defense?
What the lever to mitigation?
Where are the major gaps?
Please use it :)
I’m sharing the deck in case it’s useful to other people trying to learn about AI-biosecurity, people going through similar courses, or community organizers who want to introduce the topic to their groups.
Feel free to take it, adapt it, improve it, or use it as the basis for a discussion or short lecture.
And if you spot something that I’ve oversimplified or represented poorly, I’d be particularly interested in corrections.
The taxonomy itself is preliminary, and this deck is an even more compressed interpretation of it.
AIxBio safeguards presentation—feel free to use it
Link post
Disclaimer: This is a post written by my friend, a fellow community member, that due to her current job she can’t post out of her own forum account. If you want to reach out to her, send me a message and I’ll make the connection.
tl;dr- While participating in the BlueDot Impact’s Biosecurity course, I was requested to share a deep dive on AI-bio safeguards, so I created a slide deck on a reading that I found interesting.
Out of the readings for AIxBio unit of the Bluedot biosecurity course, one of the resources I found particularly interesting was the Frontier Model Forum’s Preliminary Taxonomy of AI-Bio Misuse Mitigations. The text tries to map the different layers of defense that could reduce the risk of someone using increasingly capable AI systems to cause severe biological harm.
I wanted a simpler way to explain the framework in a non-technical way, so I turned it into a short presentation.
Five types of interventions to mitigate AI bio misuse
The article (and the deck) organizes the mitigations into five type:
Capability limitation- trying to remove dangerous knowledge from the model entirely so it cannot help even if asked
Behavioral alignment- the model possesses dangerous knowledge but is trained to refuse sharing it
Detection & intervention- monitor how models are being used and detect or block potentially harmful interactions
Access control- restrict who can use the model, what they can access, and under what conditions
Ecosystem & societal safeguards- AI developers share threat intelligence, tools, and capabilities with governments and other actors to strengthen physical-world defenses
For each layer, I tried to compress the original brief into:
What is the defense?
What the lever to mitigation?
Where are the major gaps?
Please use it :)
I’m sharing the deck in case it’s useful to other people trying to learn about AI-biosecurity, people going through similar courses, or community organizers who want to introduce the topic to their groups.
Feel free to take it, adapt it, improve it, or use it as the basis for a discussion or short lecture.
And if you spot something that I’ve oversimplified or represented poorly, I’d be particularly interested in corrections.
The taxonomy itself is preliminary, and this deck is an even more compressed interpretation of it.