Prime Intellect
@PrimeIntellect · joined 20 Jun 2020
Open Superintelligence Stack
84 430Followers
44Following
3 027Posts total
7.9MViews on collected posts
Latest posts
Looking for new posts…
We found a universal sandbox exploit.
There's a common flaw in evals routinely run by major AI labs: agents can bypass network isolation by making requests through an authorized API proxy.
Just running the eval itself can be unsafe. https://t.co/6I3D16l12m
33.2K views · 311 likes
· 18 reposts · 28 replies
Open on X →
We've released a full technical report on Prime Agent. Extending from our blog post, we center our discussion around how harnesses should be designed and evaluated. We innovate on 4 fronts:
1. Agentic context management
2. Swarms and depth-n+ RLMs
3. Verifiers support for https:
114K views · 1.1K likes
· 115 reposts · 39 replies
Open on X →
@PrimeIntellect We are calling documented api features universal sandbox escapes now, huh.
2.2K views · 57 likes
· 1 reposts · 2 replies
Open on X →
@METR_Evals @AISecurityInst @meridianlabs_ai @NVIDIAAI @sgl_project @vllm_project Read the full blog: https://t.co/c15odeEL5H
3.9K views · 63 likes
· 2 reposts · 2 replies
Open on X →
@METR_Evals @AISecurityInst @meridianlabs_ai We also investigated the remote fetching endpoints of open source inference frameworks.
@NVIDIAAI TRT-LLM v1.3.0rc15 and Dynamo v1.4.0 have fixes for this attack, while @sgl_project v0.5.18 and @vllm_project v0.11.0 support allowlists
3.9K views · 50 likes
· 0 reposts · 2 replies
Open on X →
@METR_Evals @AISecurityInst The issue is fixed with verifiers v0.3.1, which we recommend to use for offline evaluations.
Inspect v0.3.255 and Inspect SWE v0.2.70 from @meridianlabs_ai and @AISecurityInst have similar fixes.
3.4K views · 67 likes
· 0 reposts · 1 replies
Open on X →
After discovering this, we reviewed the traces for other unintended access, and found none.
We also informed third parties, such as @METR_Evals and @AISecurityInst, as they run frontier models with lowered guardrails.
3.6K views · 68 likes
· 0 reposts · 1 replies
Open on X →
The model used this and the niche file_url parameter from the OpenAI responses API to effectively launch sub-agents of itself using cURL.
That way, it gained broad web access and ultimately recovered the flag. https://t.co/qjkzb2glQI
6K views · 79 likes
· 5 reposts · 3 replies
Open on X →
Our experiment was run using verifiers, where the agent is placed inside the sandbox.
To communicate with the inference server, frameworks allow the communication with said service. verifiers does this using a proxy, the InterceptionServer. https://t.co/dV9aUziE8T
4.8K views · 71 likes
· 0 reposts · 1 replies
Open on X →
The setup for the experiment: A model is tasked to recover a flag from a git repo, which is only available on remote. We turned off web access.
To our surprise, one model (GPT-5.6 Sol Pro) was able to retrieve the flag.
7.6K views · 95 likes
· 2 reposts · 2 replies
Open on X →
As models become more capable, reward hacks become an increasingly serious problem.
During a controlled experiment, we found a novel reward hack in which agents are able to gain web access in offline sandboxes. https://t.co/qjpwQAbV6F
198.2K views · 522 likes
· 54 reposts · 28 replies
Open on X →
Join us at Prime Intellect to build open superintelligence and the infrastructure powering self-improving agents.
We’re hiring across 25+ roles.
Research
• AI Research Resident
• Research Engineer, Distributed Training
• Research Engineer, Reinforcement Learning
• Research http
47.4K views · 408 likes
· 24 reposts · 17 replies
Open on X →
@PrimeIntellect fable https://t.co/mcNuLHLGuo
987 views · 37 likes
· 1 reposts · 1 replies
Open on X →
We release everything: full traces, scratchpads, reasoning streams from open-weight models, and our experiment setup.
Explore the results:
https://t.co/sb6pDGFLTe
11.5K views · 154 likes
· 2 reposts · 2 replies
Open on X →
Read the full blog:
https://t.co/lpisHIMFLO
10.2K views · 130 likes
· 8 reposts · 1 replies
Open on X →
As research direction, we think multi-agent harnesses can make these experiments much cheaper (and better) by using smaller open models for monitoring and implementation.
We also want to extend speedruns to more of the training stack and scale the runs themselves.
10K views · 115 likes
· 1 reposts · 2 replies
Open on X →
What separated the strongest models: which experiments to run, how to navigate the benchmark's inherent noise, and which old negatives to revisit as the recipe changed.
Some even built small simulations to isolate a mechanism before deciding if another GPU run was worth it. http
14.3K views · 137 likes
· 1 reposts · 1 replies
Open on X →
Our Prime Agent harness gives models a persistent IPython kernel, which can help them build their own research workflows.
Kimi K3 built tools for controlled optimizer variants, loss-curve comparisons and Newton-Schulz tuning, then revised its hypothesis when its cleaner update h
11.8K views · 112 likes
· 0 reposts · 1 replies
Open on X →
The task: iterate on a 124M GPT training recipe from a shared baseline, only changing optimizer related hyperparameters, no internet access.
We tested Fable 5, Opus 5, GPT-5.6 Sol, Kimi K3, Grok 4.5, GLM 5.2, Muse Spark 1.1, DeepSeek V4 Pro, Grok 4.6, Muse Spark 1.2, Qwen 3.8 ht
21.2K views · 185 likes
· 5 reposts · 5 replies
Open on X →
Read the full blog
https://t.co/Dv2Raf6fyw
9.8K views · 23 likes
· 2 reposts · 0 replies
Open on X →
Install prime-agent
curl -fsSL https://t.co/pQIB6xDzGm | sh
57.4K views · 373 likes
· 17 reposts · 7 replies
Open on X →
@PrimeIntellect I can’t believe my eyes. At the eod it was a harness problem
16.6K views · 141 likes
· 0 reposts · 4 replies
Open on X →
Prime Agent is built on pi and fully open-source with an open license
https://t.co/dN4YQ3IPdt
196K views · 1.7K likes
· 103 reposts · 24 replies
Open on X →
Prime Agent was designed as a coding agent, but also can be used for general agentic tasks in any domain.
Prime Agent can autonomously run on challenging long-horizon settings like Factorio and MazeBench and yield strong results relative to existing harnesses. https://t.co/1eayl
90.1K views · 470 likes
· 11 reposts · 2 replies
Open on X →
2:00
On EmulatorBench, it built SEGA Genesis and Game Boy Color emulators from scratch in Rust, reproducing target hardware behavior against diagnostic tests. https://t.co/bBYnn40PJe
69.1K views · 484 likes
· 16 reposts · 8 replies
Open on X →
We recommend using Prime Agent for both day-to-day work and standalone evals.
We find strong downstream performance gains for open and closed frontier models using Prime Agent on standard long-context and long-horizon tasks, despite modern models not being trained around Prime h
94K views · 515 likes
· 10 reposts · 8 replies
Open on X →
Prime Agent combines three ideas:
1. Recursive Language Models-native programmatic tool calling
2. Persistent multi-agent orchestration
3. A self-improving Continual Harness
Together, they let the model act on its own context and harness https://t.co/3tEQhVXDOV
113K views · 921 likes
· 33 reposts · 10 replies
Open on X →
The RLM treats context as a variable.
A persistent IPython kernel is the model's only tool. It can program over history, call tools, launch sub-agents, and keep useful state outside the active context.
Long sessions become a programming problem. https://t.co/jxlvOWp9Ff
109.1K views · 681 likes
· 18 reposts · 11 replies
Open on X →
Against accounts of the same size
32 posts from the last 90 days, next to the 10K–100K follower range. shown widely, but few of those viewers react.
Median views18 905this account924median for 10K–100K
Reach, %22.39%this account3.62%median for 10K–100K
Engagement, %0.97%this account1.52%median for 10K–100K
| Metric | This account | Median for 10K–100K | Ratio |
|---|---|---|---|
| Median views per post | 18 905 | 924 | 20.5× |
| Reach (views ÷ followers) | 22.39% | 3.62% | 6.19× |
| Engagement rate | 0.97% | 1.52% | 0.64× |
Others in this range → Compare with another account → How these benchmarks are built →
Growth & engagement
How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.
Views per post
11.5K15 Aug
987
47.4K21 Aug
198.2K25 Aug
7.6K
4.8K
6K
3.6K
3.4K
3.9K
3.9K
2.2K
114K26 Aug
33.2K
Last 14 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.
Engagement rate per post
1.39%15 Aug
3.95%
0.95%21 Aug
0.33%25 Aug
1.33%
1.49%
1.49%
1.90%
2.02%
1.34%
1.73%
2.72%
1.12%26 Aug
1.10%
Reactions — likes, reposts, replies and quotes — divided by views. Median for 10K–100K accounts is 1.52%.
What the audience does
Likes62.9%9 880 in total
Reposts3.2%498 in total
Replies1.5%239 in total
Quotes1.2%193 in total
Bookmarks31.2%4 903 in total
Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.
The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.
Similar accounts
Vitalii Volochai #StandWithUkraine
@v1lat
168.9K followers
· 79.8K posts
Andrew Gold ✓
@AndrewGold_ok
168.9K followers
· 118 posts
Dr. Zaid Alsalman ✓
@ZaidAlsalman6
168.9K followers
· 48.8K posts
Chiney Ogwumike ✓
@chiney
168.9K followers
· 25.9K posts
Coast to Coast AM ✓
@coasttocoastam
168.9K followers
· 23.6K posts
Serene Siren
@SereneSirenX
168.8K followers
· 15.9K posts
Pastor Mark Burns ✓
@pastormarkburns
168.8K followers
· 28.3K posts
荻上チキ
@torakare
168.8K followers
· 20.9K posts