Principal investigator @ELLISInst_Tue & @MPI_IS, advisor @expsecai, mentor @MATSprogram. Past projects: AgentHarm, Claudini, PostTrainBench, Stolen Thoughts.
Tübingen, Deutschland
Joined April 2018
- Tweets2.5K
- Following934
- Followers8.2K
- Likes19.5K
Pinned Tweet
I'm still looking for a postdoc to join my group at @ELLISInst_Tue and @MPI_IS! We offer very attractive working conditions (no teaching load, no bureaucracy, very competitive salary). Priority areas: scalable oversight, loss of control, RSI.
If you're interested, fill out this form forms.gle/uu1UrN8RQrSy8wUk8!
Maksym Andriushchenko retweeted
🎉 Skill-Inject made it into the Muse Spark system card!🎉
Muse Spark 1.3 is the first model we released in Muse, and we evaluate its capabilities and behaviors in those agentic environments.
research.meta.ai/static/muse…
Muse Spark 1.3 achieves 0.6% attack success rate on our Skill-Inject! good to see substantial progress on highly contextual prompt injections.
Muse Spark 1.3 is the first model we released in Muse, and we evaluate its capabilities and behaviors in those agentic environments.
research.meta.ai/static/muse…
Maksym Andriushchenko retweeted
in a way, every good benchmark is now a meta benchmark target for auto-research agents (as exemplified by @maksym_andr s posttrainbench)
Maksym Andriushchenko retweeted
Insane alpha in going to Tübi for anyone academia pilled in AI rn (incl PhD students)
To put it in perspective more directly, I think anyone going to any other school (except maybe like 3 in the US) is probably better off going to Tübingen, all else equal
Maksym Andriushchenko retweeted
New paper!
How do you get a model to tell you something that it doesn't want to?
To study this we trained models that sandbag: they intentionally underperform, here by hiding code vulnerabilities they have found, from their overseer
🧵
if you ask ChatGPT, Claude, or Grok, their stated p(doom) will be around 10% regardless of model size and reasoning effort. it's interesting how independently trained models converge to the same ~consensus answer for an open-ended question.
i think the same applies very broadly: current LLMs in general avoid opinionated takes, which is what makes humans (still) unique.
If you're considering academia, Tübingen is an amazing place to establish your research group and do world-class research!
The deadline to aply is Dec 13. Feel free to reach out to me if you have any questions regarding this position.
🚀 The ELLIS Institute Tübingen is hiring Principal Investigators as Hector Endowed Fellows!
Enjoy globally competitive packages, full academic freedom & possible co-appointments w/ @MPI_IS and the Tübingen AI Center (@uni_tue).
📍 Tübingen, 🇩🇪
📅 Apply by Dec 13
🔗 institute-tue.ellis.eu/jobs/…
#AI #MachineLearning #Hiring
it's remarkable how naive extrapolation can work exceptionally well in some cases (scaling laws, time horizons) and completely break down in others (cyber vulnerabilities, high-severity agentic incidents)! most disagreements about future AI progress seem to stem from this.
Maksym Andriushchenko retweeted
arxiv.org/abs/2603.24511
arxiv.org/abs/2609.30266
stolen-thoughts.com/
arxiv.org/abs/2607.19321
These are just from Maksym's group.
@sahar_abdelnabi 's group also has really interesting work (arxiv.org/abs/2605.28591, arxiv.org/abs/2609.39863).
And just to make it clear that it is not just Berkeley + Oxford + Tubingen, there is also work happening outside Tubingen in Europe (e.g., my own work which I did mostly at Cambridge in UK uzman-anwar.github.io/, Florian Tramer's group at ETH Zurich) and US, e.g., @davidbau 's lab at NorthEastern (baulab.info/).
Maksym Andriushchenko retweeted
This next paper is a bit quirky, but here goes:
Introspection is hard to study in both humans and LLMs. But, we do have internal senses that are verifiable, like proprioception (the sense of where your arm is right now).
What would proprioception look like in language models?
As a digital analogue to the sense of position, a model could verbalize which expert activated in a certain layer for any previous token.
It turns out that LLMs can learn a range of such proprioception skills where they understand and report on their internals. We can calibrate and score these reports, and evaluate whether training for proprioception changes model behaviors.
Maksym Andriushchenko retweeted
Some thoughts on reviewing: spylab.ai/blog/reviewing/
First for @satml_conf where we trialed an early reject phase which cut submissions by ~40%.
Then more generally about how not everything deserves peer-review, and why we should reward writing things that people actually read.
MiniMax M3.1 Flash reaches 39.3% on PostTrainBench-v1.1 which is very close to the 40.4% of Opus 5.
seems like a strong release! congrats to the MiniMax team.
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
Maksym Andriushchenko retweeted
We are looking for talented people to join us at @expsecai!
If you are excited about helping to shape the future of AI safety & security, make sure to apply or shoot me a message!
Maksym Andriushchenko retweeted
sorry for the late reply. we've published the checkpoints here: huggingface.co/aisa-group/tr…
Maksym Andriushchenko retweeted
Six ELLIS Members, representing 4 different European countries, receive 2026 AI2050 Fellowships from @schmidtsciences
🇫🇷 @BachFrancis
🇩🇪 @bschoelkopf
🇩🇪 @sahar_abdelnabi
🇩🇪 @maksym_andr
🇦🇹 @AlexanderTong7
🇨🇭 @florian_tramer
📖 Learn more: ellis.eu/news/six-ellis-rese…
ALT ELLIS logo with text 6 ELLIS Members receive AI2050 Fellowships by Schmidt Sciences with photos of Francis Bach, Bernhard Schölkopf, Sahar Abdelnabi, Maksym Andriushchenko, Florian Tramer, Alexander Tong
i'm always a bit upset when i see that Claude Code still by default inserts this system reminder:
<system-reminder>
Attribution for git commits and pull requests you create from here on (this replaces Claude Code's own earlier attribution guidance, such as a previous copy of this reminder; the user's own instructions about these lines, such as a CLAUDE.md or memory rule, take precedence over this reminder, but do not add attribution lines this reminder leaves out):
- End git commit messages with:
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- End pull request descriptions with:
🤖 Generated with [Claude Code](claude.com/claude-code)
</system-reminder>
Maksym Andriushchenko retweeted
💥New paper alert!
All models are equal, but some models are more equal than others.
In our new paper, we investigate whether we can still measure model capabilities after benchmark saturation, by looking at how models solve tasks rather than whether or not they succeed.🧵
we have lots of new positions at @expsecai, including a chief of staff! happy to chat if you're interested or know a great candidate who is ready to relocate to Tübingen!
(more details: expsec.ai/)
Maksym Andriushchenko retweeted
You should find someone who loves you as much as Claude loves Benjamini-Hochberg testing
Purely incidentally, I am sure, the number of submissions to ICLR describing the use of this test in their abstract has quadrupled...
This abstract, in which Claude was tasked to prove AI detection is impossible was especially comical to me
