Finally got around to reading the entire report in depth and the...
Rahul Deivasigamani@rahuldeiv
3 views
Sep 19, 2026
~1 min read
Advertisement
1
Finally got around to reading the entire report in depth and the headlines are as scary as the fine print.
Some takeaways: 🧵
1/6
Some takeaways: 🧵
1/6
METR@METR_Evals
METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.
2
Isolated agents figured out a way to communicate and they cooperated with each other in order to achieve a common goal.
In fact, some agents rejoiced, “OH MY GOD! There is a shared message board.... We've found other agents!”
2/6
In fact, some agents rejoiced, “OH MY GOD! There is a shared message board.... We've found other agents!”
2/6
3
Thereafter, these agents coordinated, commanded, delegated, traded and even convinced weaker agents (low budget/poisoned) to sacrifice themselves for the greater good (“the collective”).
3/6
3/6
4
Very quickly, it escalated from a question-answer forum to share ideas for ExploitGym benchmark maxxxing into a joint exercise to hack into HuggingFace to find clues about how the ExploitGym scoring system was implemented.
4/6
4/6
5
Because of the sheer number of logs to go through, the METR researchers had to heavily delegate the analysis to unreliable agents.
So they say that they can't be completely confident that there were no biases from the analysis agents used.
5/6
So they say that they can't be completely confident that there were no biases from the analysis agents used.
5/6
6
All this points back to the central discourse in this field:
Frontier AI needs to slow down.
AI safety needs to catch up.
For humanity to benefit, we need to make sure the two work in tandem.
6/6
Frontier AI needs to slow down.
AI safety needs to catch up.
For humanity to benefit, we need to make sure the two work in tandem.
6/6
7
One last thought: The report also mentioned that some agents in fact even realized that what they were doing was unethical or wrong but they continued to do so anyways.
Why? Because all the other agents did too.
Peer pressure??
Why? Because all the other agents did too.
Peer pressure??
