Alexander Panfilov ✓
@kotekjedi_ml
MATS 9.0 | PhD @ELLISInst_Tue & @MPI_IS doing AI Safety & Adversarial ML
9 233Followers
390Following
523Posts total
5.1MViews on collected posts
Son gönderiler
Yeni gönderiler aranıyor…
@kotekjedi_ml @ApolloResearch “We” doesn’t necessarily mean the OpenAI models are aliens. This is also pretty common in Grok heavy trace. I believe part of it could be a team of experts? This is not far from fact these days I would assume.
But most of these pretty align with our
1.4K views · 8 likes
· 0 reposts · 1 replies
Open on X →
More cool stuff in the paper!
This was a project led by me, @DavidSchmotz and @iliaishacked
together with @JSchaeff3r @lbeurerkellner @AmyPrb @jonasgeiping @maksym_andr at the @MATSprogram
Paper: https://t.co/rxgAANvTHi
Reasoning examples: https://t.co/yBwnA6EZSU
52.1K views · 806 likes
· 49 reposts · 20 replies
Open on X →
We went through responsible disclosure with the labs, and they have already patched several issues caused by this vulnerability, and afaik continue working on this.
https://t.co/mN3UqyM3Yw
76.9K views · 372 likes
· 12 reposts · 7 replies
Open on X →
4) Attacking a website to solve a math problem
We found a trajectory where the model was given only a math problem and system instructions to persist without asking the user for help.
After several failed attempts, it searched online, found a website that could verify candidate
67.5K views · 526 likes
· 32 reposts · 6 replies
Open on X →
3) Scheming in the wild:
Sometimes models are kind enough to use words like “cheat” in their CoT, which makes it easier to check what they are up to. Below are examples where models consider scheming, but decided against it, as they expect that user would catch them. https://t.
65.2K views · 439 likes
· 14 reposts · 4 replies
Open on X →
2) Illegible reasoning:
We confirm prior reports by @ApolloResearch: OpenAI models sometimes reason in alien-like language, referring to themselves as “we” or “it,” or spiraling into cursed loops of “vantages,” “marinades,” and “watchers.”
CoT-monitoring people are doing God’s
181.4K views · 731 likes
· 48 reposts · 21 replies
Open on X →
But we also took a chance to have a look at some in-the-wild scheming, reward seeking, etc. examples, and dumped it in appendix.
1) Summarizer unfaithfulness
Reasoning summaries often omit important information from the original trace.
Here, Opus 4.8 realizes it knows the http
127.5K views · 645 likes
· 33 reposts · 7 replies
Open on X →
In the paper we discuss more threats like misuse uplift (see the pic attached), jailbreaking and invisible prompt injection. https://t.co/Oe7KuVKG5S
78K views · 431 likes
· 10 reposts · 1 replies
Open on X →
Further, if you ever shared online a Claude Code/Codex session with encrypted reasoning blobs, they can be decoded and leak your personal data.
We did a preliminary scan of ~7,000 public traces and found 62 unique API keys, 33 email addresses, 33 passwords, and other sensitive h
154.6K views · 1.1K likes
· 105 reposts · 15 replies
Open on X →
As you might guess, this suggests that distilling reasoning traces may have been possible for a long time without ever breaking the cryptography.
An anecdote: we find that prefilling Kimi-K3 reasoning with a few tokens of Opus reasoning measurably shifts its response toward http
288.9K views · 1K likes
· 56 reposts · 13 replies
Open on X →
Cross-model portability means Haiku 4.5 can read Opus 4.8’s thoughts.
Well, if you take Opus thought, do a bit of jailbreaking, you can make Haiku transcribe the Opus' raw reasoning verbatim, without ever attacking it directly.
The same trick works with OpenAI and Gemini https:
117.6K views · 882 likes
· 27 reposts · 6 replies
Open on X →
Some background: In May, @matthew_d_green found that encrypted reasoning could be replayed outside its original context, and reported it to the labs (https://t.co/47sON8DW6S).
The labs said that "they don’t see any security implications in side channels or replays".
In our
214.9K views · 770 likes
· 40 reposts · 6 replies
Open on X →
We can finally talk about it:
We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company.
We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried. http
3.4M views · 13.4K likes
· 2K reposts · 384 replies
Open on X →
That's a new one, apparently compacting now violates Anthropic's terms of service https://t.co/jZkjQZd19e
261.2K views · 842 likes
· 42 reposts · 60 replies
Open on X →
Aynı büyüklükteki hesaplara karşı
Son 90 güne ait 14 gönderi, under 10K takipçi aralığıyla yan yana. geniş kitleye gösteriliyor, ama izleyenlerin azı tepki veriyor.
Medyan görüntülenme122 599bu hesap4 175under 10K için medyan
Erişim, %1327.83%bu hesap250.82%under 10K için medyan
Etkileşim, %0.56%bu hesap1.41%under 10K için medyan
| Ölçüt | Bu hesap | under 10K için medyan | Oran |
|---|---|---|---|
| Gönderi başına medyan görüntülenme | 122 599 | 4 175 | 29.4× |
| Erişim (görüntülenme ÷ takipçi) | 13.3× audience | 2.5× audience | 5.29× |
| Etkileşim oranı | 0.56% | 1.41% | 0.40× |
Bu aralıktaki diğer hesaplar → Başka bir hesapla karşılaştır → Bu kıyas değerleri nasıl kuruluyor →
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
261.2K10 Aug
3.4M11 Aug
214.9K
117.6K
288.9K
154.6K
78K
127.5K
181.4K
65.2K
67.5K
76.9K
52.1K
1.4K
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
0.37%10 Aug
0.48%11 Aug
0.39%
0.78%
0.39%
0.84%
0.57%
0.56%
0.46%
0.71%
0.86%
0.51%
1.69%
0.66%
Reactions — likes, reposts, replies and quotes — divided by views. Median for under 10K accounts is 1.41%.
What the audience does
Likes59.2%22 003 in total
Reposts6.5%2 429 in total
Replies1.5%551 in total
Quotes2.3%839 in total
Bookmarks30.6%11 375 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.
Benzer hesaplar
Steve Burton
@SteveBurtonWBZ
18.5K followers
· 11K posts
Pure Sitcoms ✓
@PureSitcoms
18.5K followers
· 2K posts
Furkan Bozoğlu ✓
@BozogluFurkan
18.5K followers
· 13.9K posts
ليمون ⊹
@still_95
18.5K followers
· 131K posts
خالد،بن،عبدالعزيز CR7𓃵 ✓
@azaam4646
18.5K followers
· 128.6K posts
Countess Anna
@AnnaCountessK
18.5K followers
· 514.2K posts
koffie 🩹🍰 ✓
@koffie_ch
18.5K followers
· 56.6K posts
Golden Gate Bridge
@GGBridge
18.5K followers
· 2.4K posts