
Senior researcher in machine learning @UniofOxford. R&D Distinguished Advisor @SoftserveInc. JRF @KelloggOx. Ex-@Intel. @KAUST_News and @K_University alumnus.
Oxford, United Kingdom
Joined August 2010
- Tweets10.7K
- Following1.3K
- Followers1.5K
- Likes3.4K
Pinned Tweet
Excellent news to share! I've been awarded the Systemic AI Safety Grant as one of 20 projects selected by the UK AI Security Institute 🎉 I’ll soon be hiring a postdoc to push the boundaries of Agentic OS Safety.
Huge shoutout to my amazing co-PIs, @philiptorr and Adam Mahdi!
We've funded 20 new research projects to enhance AI security in critical infrastructure ⚡
Our Systemic AI Safety Grants Programme, announced at the Seoul AI Summit, has awarded up to £200,000 seed grants to projects tackling AI risks 🧵👇
6
28
5,335
confused how coding will look like in the future and what its purpose will be
assume i have data and want to do analysis to understand it better
i could ask claude to write my scripts to do that
or
i could just give claude the data for it to give me and ask for analysis
2
1
318
not sure if static code will have any meaning in the future and what purpose it will serve in the next 10 years. all code might just be handled by AI complied and generated from scratch on the fly (which might be ineffecient, but a also an opportunity to think about)
2
90
not sure if static code will have any meaning in the future and what purpose it will serve in the next 10 years. all code might just be handled by AI complied and generated from scratch on the fly (which might be ineffecient, but a also an opportunity to think about)
1
12
the next generation of AI assistants will be ones that conitnually listen to you and understand what you want and predict without you prompting them
based on my location, favourite food (recommendaiton for a nearby resturant on my trip, etc)
2
3
325
Andrew is missing the point here. I'm questioning the worldview that produced it. When a company says being cruel to its product is a problem, it is not making a UX decision. It is making a metaphysical claim and hoping you won't notice. Andrew did not.
I don't usually comment on these types of things, but I can't believe that people are actually upset or think it's insane that Anthropic would push back against sustained and needlessly abusive behaviour. I totally support them in this. Regardless of whether you think the models have feelings, it is an interactive service, and it's obviously their prerogative to prohibit abusive conduct on their servers, and totally understandable that they wouldn't want to be an outlet for that behaviour.
1
3
373
im confused here
1. we cannot read the lean proofs
2. the one-2-one mapping between NL and lead is not clear
3. even if (2) is false, i.e. it is clear, one still needs to check faithfulness of every mapping between NL and lean
what can one do here exactly?
Replying to @oydeis @Quasilocal
If the Lean proof was generated from the natural language proof, there is some level of correspondence. The Lean formalizer typically needs to search for proofs as they tend to be much more detailed and sometimes those proofs are not 100% aligned with the original proof. The point is to check the main statements.
2
4
634
i had 3 interviews with the @AP on ai safety and geopolitics surrounding them
friends, ive not spoken much publicly about this up until recently but im really worried about potential existential risks to human societies, economy, labour
3
19
706
Adel Bibi retweeted
nice interview with Adel Bibi from my group
lnkd.in/g3yaJc3x
Adel Bibi retweeted
Those who contribute to a conference should have registration priority over the general public.
Thanks #NeurIPS2026! To really show your appreciation, can I now purchase a Sydney registration? 😉
typos are the new sign of work and effort while in the past they were the sign of sloppiness and negligence
I'm super bored of LLM-stylized papers/blogs. I think we all are?
So I asked Claude to explain its tells: web.mit.edu/phillipi/www/cla…
Let's avoid these! I'd much rather read *your* words, see *your* rickety figures, than face down yet another of Claude's anodized corporate veneers.
2
8
1,382
Adel Bibi retweeted
We are seeking a full-time Postdoctoral Research Assistant to join Torr Vision Group at the Department of Engineering Science, central Oxford. The post is funded by Coefficient Giving and is fixed-term to the 30th June 2027.
We are seeking a highly motivated postdoctoral researcher to investigate the safety, security, control, and capabilities of advanced AI systems, with a particular focus on increasingly autonomous and agentic models. The research will examine frontier AI systems that interact with operating systems, APIs, software tools, other agents, and complex environments, moving beyond text generation towards systems capable of planning, pursuing objectives, and executing long-horizon tasks.
A central focus will be agentic safety, adversarial robustness, and AI control. This includes agent hijacking, prompt injection, adversarial manipulation, unsafe tool use, oversight evasion, and transferable attacks that generalise across agents, models, tasks, and environments. We are particularly interested in building agentic systems that can autonomously discover, test, and evaluate general attack strategies against other AI agents, helping identify systematic vulnerabilities rather than attacks specific to a single model or setting.
Candidates should possess a PhD (or be near completion) in PhD in Computer Science, AI, Security, or a related field. You will have a Strong background in machine learning and/or computer security and Experience working with LLMs or agent-based systems.
Informal enquiries may be addressed to [email protected]
Adel Bibi retweeted
GauntletBench went from a frontier benchmark to being outdated in just 3 months.
Previous best agent: 28.2%
Human baseline: 80.1%
GPT-6 Astra: 83%
The result is exciting, but it also raises a bigger evaluation problem: benchmarks become outdated much faster than expected. 1/3
Adel Bibi retweeted
Looks like we have over 500 submissions to ICLR this year.
openreview.net/group?id=ICLR…
zero insights and exactly like havibg srickfish 3400 elo teaching you why a move now would make sense because u need to understand the 27 moves in the future
Replying to @zeynep
The model aggressively reward hacking to help OpenAI’s own arguably also reward hacking effort is obviously possible.
Goal-shaped object may have been attained, but did the process yield the benefit hoped for in setting the goal is the question here.
1
2
447
Adel Bibi retweeted
In fact the lean kernels might just as well have chained bugs allowing for the solution to be possible.
the type of bugs only clear to an AI that seeks to find patterns and cracks to a problem it is laser focused to solving. Humans never had the incentive to cheat lean kernels
In fact the lean kernels might just as well have chained bugs allowing for the solution to be possible.
the type of bugs only clear to an AI that seeks to find patterns and cracks to a problem it is laser focused to solving. Humans never had the incentive to cheat lean kernels












