AI researcher, 0day enthusiast @xchglabs
nyc
Joined March 2022
- Tweets110
- Following35
- Followers56
- Likes180
Royce Moon retweeted
🧵[1/N] Are large language models more creative than humans at combining ideas? 🤔
Combinatorial creativity involves combining familiar ideas in unfamiliar ways, and is central to scientific discovery and technological innovation.
In fact, breakthrough scientific work is often associated with more distant combinations of methods, techniques, and concepts.
Here, we introduce Kombine, a new benchmark of combinatorial creativity for LLMs.
TLDR:
- LLMs are more combinatorially creative than the average human.
- However, conceptual blending is the hardest task for LLMs, and even frontier models struggle.
- LLMs also tend to homogenize, inventing concepts with the same properties across separate models.
- Moreover, humans and LLMs combine in distinct regions of semantic space.
- As a result, we anticipate greater conceptual diversity can be unlocked through human-AI collaboration. 🗣️
Thread below🧵⬇️
Royce Moon retweeted
Today I’m excited to announce the public release of Scend
Over the last year, we’ve been working with private equity firms, investment banks, consulting firms, and search funds to power their market research and deal sourcing
M&A currently has a data problem
More than ever, PE is focused on acquiring and modernizing legacy businesses. The problem is that many of these firms have little to no data footprint. You won’t find them in legacy financial databases, and their founders often don’t have LinkedIn. So your ideal acquisition target will be in the long-tail of companies in the space; but it won’t exist in traditional finance databases that firms subscribe to.
Scend takes a different approach: it isn’t a database at all
Give Scend a prompt and it can generate any industry-specific dataset or research you can imagine. Each value is fetched and cited by an agent - so if you wanted to estimate the revenue of a dental clinic for example, it might:
1. find that clinic’s website
2. find the page listing out the number of staff members
3. multiply (headcount) x (industry multiple) to estimate revenue
4. then cite its sources and give its methodology
All to produce a single value in your dataset.
Our mission in building Scend is to give private-market investors the heavy duty research and data infrastructure that public market investors have had for decades
See what we built below 👇
We’re hiring the best in hardware reversing for an exciting new project!
Something that has never been done before 👀
We’re hiring a Hardware Reverse Engineer at Xchg Labs.
3-month, part-time contract. Remote is fine, but DMV researchers are preferred.
You’ll build hardware testing platforms, extract firmware from extremely difficult systems, and document the process.
No degree required. Bonuses included. You’ll also receive credit for approved work and can publish a technical blog with us afterward.
Apply: linkedin.com/jobs/view/44496…
Royce Moon retweeted
What happens when you point one of the most interesting AI models available at the Linux Bluetooth stack?
We used @autopoiesislab Colombus-1 model to help uncover vulnerabilities in BlueZ.
Full research from Xchg Labs:
xchglabs.com/blog/bluez-zero…
Congrats on the release! Happy to partner with @autopoiesislab using their new model Colombus-1 to do some cool vuln research on BlueZ.
We found a zero-click full chain RCE 0day! Blog post incoming from @xchglabs too 👀
Reinforcement learning rewards convergence, and convergence is the opposite of discovery.
We built the world’s first judgment model from the decisions papers erase: what to pursue, challenge, abandon, and revisit. Columbus-1 autonomously found a zero-click Bluetooth RCE and designed a 10-foot, self-landing rocket.
Royce Moon retweeted
We pointed Columbus-1 at BlueZ because mature wireless stacks are where shallow search fails.
It found 8 confirmed vulnerabilities, including a nation-state-tier zero-click full-chain RCE: no click, pairing, or interaction. @XchgLabs and Enclave Intelligence (@compulsi0n) independently reproduced and confirmed the findings.
Royce Moon retweeted
First Qemu escape, kinda nervous...
xchglabs.com/blog/escaping-q…
a bit late of a notice but happy to share a paper i've been cooking up with @lrvarshney was accepted to the #ICML2026 Second Workshop on Agents in the Wild!
we've been working on AI safety guarantees independent of alignment (arxiv link in comments).
paper link: arxiv.org/abs/2605.09045
Royce Moon retweeted
One malicious SMTP server and one EHLO AUTH line.
RCE on the lwIP client!
We got 255-byte tx_buf, ~64 KiB of attacker bytes, and zero auth. Demo below.
Patched as Savannah #68313. Writeup is at:
xchglabs.com/blog/lwip-smtp-…
#infosec #RCE #IoT #embedded #vulnresearch #smtp
Royce Moon retweeted
Two new writeups dropped!
On lwIP SMTP: xchglabs.com/blog/lwip-smtp-…
On KVM dirty-ring: xchglabs.com/blog/kvm-dirty-…
#infosec #lwIP #KVM #linux #IoT
Royce Moon retweeted
Hmmm... it seems like Mythos and Codex-5.5-Cyber haven't found *all* of the exploitable vulnerabilities just yet?
It will be especially interesting to see the results in the browsers category this year, where Chrome and Firefox have been able to apply LLMs to fix what they can.
Pwn2Own hit max capacity for the first time in history. We unfortunately couldn’t get in. We submit our research to vendors. Look to our blog for technical breakdowns coming soon!
#cybersecurity #0day #reverseengineering #exploitdev #infosec #xchglabs #pwn2own
Royce Moon retweeted
Pre-auth remote heap buffer overflow in dnsmasq is wild: xchglabs.com/blog/dnsmasq-fi…
Our new paper introduces containment verification where we formally verify interactions between agents and the world via havoc oracle semantics, allowing for capability-agnostic, universal formal safety guarantees.
We hope this can make a tangible path to trusted, safe AI!
Royce Moon has done really incredible work on a containment layer for agentic AI. It ensures AI safety without having to dig into the internals or alignments of AI models, by formally verifying interactions with the world. Works whether considering superintelligence, super-duper intelligence, or anything beyond. arxiv.org/abs/2605.09045
5 CVEs in dnsmasq! more to come
dnsmasq powers massive amounts of networking infrastructure worldwide.
We found 6 0days. 5 were accredited to xchglabs and assigned CVEs.
Our first technical writeup is now live:
xchglabs.com/blog/dnsmasq-fi…
#dnsmasq #0day #cybersecurity #xchglabs
Stay tuned for everything from @xchglabs 👀
‼️🚨 Pwn2Own Berlin 2026 just hit a wall. For the first time in 19-years, ZDI rejected dozens of working zero-day RCE submissions because organizers ran out of contest slots.
Rejected hackers are now going public with PoC demos and direct vendor disclosures, breaking Pwn2Own's usual secrecy.
▪️ AI surfaces a massive wave of 0-day RCEs.
▪️ Submissions overwhelm ZDI past max capacity.
▪️ Slots run out. Researchers with working chains get rejected.
▪️ "Revenge disclosures" begin. ← we are here.
Confirmed casualties so far:
▪️ @xchglabs : 86 vulnerabilities prepared (PyTorch, NVIDIA, Linux KVM, Oracle, Docker, Ollama, Chroma, LiteLLM, llama.cpp). All rejected. Now reporting directly to vendors with writeups dropping as patches land.
▪️ @ggwhyp : full-chain Firefox RCE on Windows. Rejected. Publicly demoed (HTML page → cmd.exe → calc.exe). Responsibly disclosed to Mozilla.
▪️ @yunsu_dev : working RCE chain, rejected. Submitting elsewhere.
▪️ @ryotkak : tried to register for 3+ weeks. ZDI confirmed "at maximum capacity, can't add extra contest days." Considered canceling flight and hotel.
▪️ @anzuukino2802 : Claude Code RCE PoC. Rejected.
▪️ @desckimh : 0-day RCEs in Ollama and LM Studio. Rejected.
Reported impact: a community-estimated 150+ researchers tried to register. Accepted contestants are now being warned about collisions. Rejected vulnerabilities going to bug bounty programs may trigger pre-event patches that invalidate the work of those who got in.
ZDI has not publicly addressed the capacity issue. The event still runs May 14-16 in Berlin.
Royce Moon retweeted
We are please to announce this year's best papers!
The best paper awards of this year’s conference was sponsored by the Artificial Intelligence Journal (AIJ).
#iccc25 #computationalcreativity #creativity #campinas #brazil #computerscience #computerengineering #computers
another banger
We’re excited to share that our recent paper, “Transformational Creativity in Science: A Graphical Theory,” co-authored by @samschapiro, @jonahablack, and @lrvarshney received the Best Short Paper Award at the 16th International Conference on Computational Creativity, organized by the @ACC_assoc.
This work opens new avenues for developing AI systems capable of achieving breakthrough insights, mirroring the cognitive leaps seen in historical scientific revolutions such as Einstein’s theory of relativity, the heliocentric model, and the advent of non-Euclidean geometry.
We’re grateful to ICCC for this recognition and look forward to continue developing a future where AI can contribute to robust, interpretable, and even transformative scientific discoveries.
Paper link: arxiv.org/abs/2504.18687
#creativity #artificialintelligence #science #innovation
