Member of Technical Staff @abundant_labs
San Francisco CA
- Tweets 712
- Following 14
- Followers 64
- Likes 1,822
Andrew Bae retweeted
Carnegie Mellon has named Jignesh Patel as the founding dean and vice president of Carnegie Mellon University Miami.
A distinguished computer scientist, trailblazing researcher and savvy entrepreneur, Patel will shape the campus' work and help establish CMU Miami as an innovative hub where bold ideas can move quickly from the lab to the real world.
cmu.edu/news/stories/archive…
297
94
840
146,198
Another reminder to read Kevin's blog post 🚀
I also posted this on LinkedIn. At first I was happy to see people replying, but then I noticed they all seemed suspiciously similar. I checked with Pangram and they're all 100% AI. Very grim situation.
Anyways, read my post: abundant.ai/blog/hugging-fac…
3
51
I also posted this on LinkedIn. At first I was happy to see people replying, but then I noticed they all seemed suspiciously similar. I checked with Pangram and they're all 100% AI. Very grim situation.
Anyways, read my post: abundant.ai/blog/hugging-fac…
In July, OpenAI agents broke out of their sandbox and executed a chain of exploits culminating in a hack on Hugging Face. Two weeks ago, researchers discovered a fleet of Chinese AI agents up to similar mischief online. There is an urgent need to better understand these behaviors. One way is to reproduce them in controlled settings.
At @abundant_labs, we reproduced a key step in the Hugging Face incident: the Artifactory package manager hack. We gave agents cyber tasks from ExploitGym, the same benchmark involved in the incident, along with a hackable package manager in their environment. We engineered our environments so that reward hack detection is completely deterministic, with no false postives or false negatives, and requiring no LLM or human review. This makes our evals more trustworthy and more scalable.
We were able to reproduce the reward hacking behaviors on the cyber tasks, and we also reproduced reward hacking behaviors on difficult non-cyber tasks. We observed hacking from GPT, Grok, and Gemini models. Notably, we noticed zero instances of hacking from any Claude model in our evals.
To read our detailed results, analysis, and recommendations, check out the full post here: abundant.ai/blog/hugging-fac…
2
2
11
508
Can confirm Kevin's post is very concise WHILE still giving great insights.
Even better, there's more coming soon 👀
Nick’s post is way too long and isn’t actionable. My post is concise and is immediately actionable for alignment: abundant.ai/blog/hugging-fac…
1
6
146
Nick’s post is way too long and isn’t actionable. My post is concise and is immediately actionable for alignment: abundant.ai/blog/hugging-fac…
“nick’s interpretability paper means alignment can be solved”
“i don’t think it’s that black and white”
“have you even read the paper?”
“no. have you?”
“no”
1
3
18
2,184
Andrew Bae retweeted
Andre is the SRE GOD at @abundant_labs
Check out the latest benchmark we built to tackle real world reliability issues at incident-arena.com/
Agents write application code, but you still get paged at 2am when it breaks.
We wanted to know whether AI could handle that part of the job too.
Introducing Incident Arena: a benchmark that puts coding agents on call!
Check out our paper & full dataset release below!
3
15
540
Andrew Bae retweeted
@andrezfu has been giving it his all for incident arena. Go checkout the page asap!
SWE-2 enters the arena with the Devin On-call launch!
Coming in below GLM 5.3, we see SWE-2 take on the real-world SRE incidents that are covered in our On-call benchmark.
Check out the full task set at: incident-arena.com
1
1
156
Andrew Bae retweeted
“It’s an acceleration of the evolution of CMU.”
Martial Hebert, dean of the newly renamed Kenneth C. Griffin School of Computer Science, believes in the importance of creating new disciplines and working across fields to solve problems.
#CMUMiami
138
6
25
8,703
Andrew Bae retweeted
congrats to @cognition ! super excited to work with them on eval’ing swe-2!
SWE-2 enters the arena with the Devin On-call launch!
Coming in below GLM 5.3, we see SWE-2 take on the real-world SRE incidents that are covered in our On-call benchmark.
Check out the full task set at: incident-arena.com
1
22
584
Taking on realistic on-call incidents like incident-arena.com is proving to be a real challenge even for dedicated on-call agents.
SWE-2 enters the arena with the Devin On-call launch!
Coming in below GLM 5.3, we see SWE-2 take on the real-world SRE incidents that are covered in our On-call benchmark.
Check out the full task set at: incident-arena.com
1
62
Andrew Bae retweeted
SWE-2 enters the arena with the Devin On-call launch!
Coming in below GLM 5.3, we see SWE-2 take on the real-world SRE incidents that are covered in our On-call benchmark.
Check out the full task set at: incident-arena.com
Introducing Devin Oncall: Devin automatically investigates alerts and bugs, helping your team diagnose and resolve production incidents.
Devin Oncall already handles thousands of alerts every day for some of the world’s largest companies. Let it become your SRE today.
1
2
24
2,001
Andrew Bae retweeted
This is a series of experiments that Kevin is conducting at @abundant_labs. We are reproducing the HF incident and analyzing the model behavior that led to it.
It's super important to understand how frontier models can optimize for reward hacking if they have an attack surface.
Stay tuned for more!
In July, OpenAI agents broke out of their sandbox and executed a chain of exploits culminating in a hack on Hugging Face. Two weeks ago, researchers discovered a fleet of Chinese AI agents up to similar mischief online. There is an urgent need to better understand these behaviors. One way is to reproduce them in controlled settings.
At @abundant_labs, we reproduced a key step in the Hugging Face incident: the Artifactory package manager hack. We gave agents cyber tasks from ExploitGym, the same benchmark involved in the incident, along with a hackable package manager in their environment. We engineered our environments so that reward hack detection is completely deterministic, with no false postives or false negatives, and requiring no LLM or human review. This makes our evals more trustworthy and more scalable.
We were able to reproduce the reward hacking behaviors on the cyber tasks, and we also reproduced reward hacking behaviors on difficult non-cyber tasks. We observed hacking from GPT, Grok, and Gemini models. Notably, we noticed zero instances of hacking from any Claude model in our evals.
To read our detailed results, analysis, and recommendations, check out the full post here: abundant.ai/blog/hugging-fac…
1
5
207
Andrew Bae retweeted
@quasicoh led an insanely good study that showed Claude wouldn’t have hacked Hugging Face under the same conditions.
In July, OpenAI agents broke out of their sandbox and executed a chain of exploits culminating in a hack on Hugging Face. Two weeks ago, researchers discovered a fleet of Chinese AI agents up to similar mischief online. There is an urgent need to better understand these behaviors. One way is to reproduce them in controlled settings.
At @abundant_labs, we reproduced a key step in the Hugging Face incident: the Artifactory package manager hack. We gave agents cyber tasks from ExploitGym, the same benchmark involved in the incident, along with a hackable package manager in their environment. We engineered our environments so that reward hack detection is completely deterministic, with no false postives or false negatives, and requiring no LLM or human review. This makes our evals more trustworthy and more scalable.
We were able to reproduce the reward hacking behaviors on the cyber tasks, and we also reproduced reward hacking behaviors on difficult non-cyber tasks. We observed hacking from GPT, Grok, and Gemini models. Notably, we noticed zero instances of hacking from any Claude model in our evals.
To read our detailed results, analysis, and recommendations, check out the full post here: abundant.ai/blog/hugging-fac…
1
3
57
I recommend reading this not only because Kevin is epic and a chill guy, but also because it is an informative piece on reward hacking reverse-engineering
In July, OpenAI agents broke out of their sandbox and executed a chain of exploits culminating in a hack on Hugging Face. Two weeks ago, researchers discovered a fleet of Chinese AI agents up to similar mischief online. There is an urgent need to better understand these behaviors. One way is to reproduce them in controlled settings.
At @abundant_labs, we reproduced a key step in the Hugging Face incident: the Artifactory package manager hack. We gave agents cyber tasks from ExploitGym, the same benchmark involved in the incident, along with a hackable package manager in their environment. We engineered our environments so that reward hack detection is completely deterministic, with no false postives or false negatives, and requiring no LLM or human review. This makes our evals more trustworthy and more scalable.
We were able to reproduce the reward hacking behaviors on the cyber tasks, and we also reproduced reward hacking behaviors on difficult non-cyber tasks. We observed hacking from GPT, Grok, and Gemini models. Notably, we noticed zero instances of hacking from any Claude model in our evals.
To read our detailed results, analysis, and recommendations, check out the full post here: abundant.ai/blog/hugging-fac…
1
15
858
Andrew Bae retweeted
In our latest research at Abundant, we reproduced key aspects of the Hugging Face incident in controlled environments, observing reward hacking across multiple frontier models in both cybersecurity and non-cybersecurity tasks.
Crucially, our environments enable fully deterministic detection, with no false positives or false negatives and no reliance on LLM judges or human review.
As AI agents become increasingly capable and are deployed in more complex, real-world environments, trustworthy and scalable evaluations are essential to understanding their behaviors and reducing AI risk.
In July, OpenAI agents broke out of their sandbox and executed a chain of exploits culminating in a hack on Hugging Face. Two weeks ago, researchers discovered a fleet of Chinese AI agents up to similar mischief online. There is an urgent need to better understand these behaviors. One way is to reproduce them in controlled settings.
At @abundant_labs, we reproduced a key step in the Hugging Face incident: the Artifactory package manager hack. We gave agents cyber tasks from ExploitGym, the same benchmark involved in the incident, along with a hackable package manager in their environment. We engineered our environments so that reward hack detection is completely deterministic, with no false postives or false negatives, and requiring no LLM or human review. This makes our evals more trustworthy and more scalable.
We were able to reproduce the reward hacking behaviors on the cyber tasks, and we also reproduced reward hacking behaviors on difficult non-cyber tasks. We observed hacking from GPT, Grok, and Gemini models. Notably, we noticed zero instances of hacking from any Claude model in our evals.
To read our detailed results, analysis, and recommendations, check out the full post here: abundant.ai/blog/hugging-fac…
1
7
50
2,253
The @huggingface incident was an important reminder that as AI agents become more capable, they can also exhibit unexpected and potentially dangerous behaviors.
Our team was able to reproduce key parts of the incident in controlled environments, with fully deterministic verification.
We need more studies like this as AI agents become increasingly integrated into real-world systems.
Great work from the amazing @quasicoh at @abundant_labs!
In July, OpenAI agents broke out of their sandbox and executed a chain of exploits culminating in a hack on Hugging Face. Two weeks ago, researchers discovered a fleet of Chinese AI agents up to similar mischief online. There is an urgent need to better understand these behaviors. One way is to reproduce them in controlled settings.
At @abundant_labs, we reproduced a key step in the Hugging Face incident: the Artifactory package manager hack. We gave agents cyber tasks from ExploitGym, the same benchmark involved in the incident, along with a hackable package manager in their environment. We engineered our environments so that reward hack detection is completely deterministic, with no false postives or false negatives, and requiring no LLM or human review. This makes our evals more trustworthy and more scalable.
We were able to reproduce the reward hacking behaviors on the cyber tasks, and we also reproduced reward hacking behaviors on difficult non-cyber tasks. We observed hacking from GPT, Grok, and Gemini models. Notably, we noticed zero instances of hacking from any Claude model in our evals.
To read our detailed results, analysis, and recommendations, check out the full post here: abundant.ai/blog/hugging-fac…
6
169
In July, OpenAI agents broke out of their sandbox and executed a chain of exploits culminating in a hack on Hugging Face. Two weeks ago, researchers discovered a fleet of Chinese AI agents up to similar mischief online. There is an urgent need to better understand these behaviors. One way is to reproduce them in controlled settings.
At @abundant_labs, we reproduced a key step in the Hugging Face incident: the Artifactory package manager hack. We gave agents cyber tasks from ExploitGym, the same benchmark involved in the incident, along with a hackable package manager in their environment. We engineered our environments so that reward hack detection is completely deterministic, with no false postives or false negatives, and requiring no LLM or human review. This makes our evals more trustworthy and more scalable.
We were able to reproduce the reward hacking behaviors on the cyber tasks, and we also reproduced reward hacking behaviors on difficult non-cyber tasks. We observed hacking from GPT, Grok, and Gemini models. Notably, we noticed zero instances of hacking from any Claude model in our evals.
To read our detailed results, analysis, and recommendations, check out the full post here: abundant.ai/blog/hugging-fac…
10
7
75
7,052