Travel Support for NEURIPS PARIS 2026
Presenting SPEC-GAP and Connecting with AI Safety Researchers at NeurIPS 2026
Travel Support for NEURIPS PARIS 2026
I’m excited that our work on hidden attacks in multi-agent AI systems was accepted for presentation at the New in ML Workshop at NeurIPS 2026 in Paris. We are a new, self-funded team, so this would be the first time anyone from our group attends the conference. I need help getting there. What I can do depends on how much I raise:
$3,465 is the minimum to get me to Paris to present the poster: the flight, registration, poster printing, and hotel.
$4,500 covers my full trip, adding meals, local transport, and a buffer.
$5,500 adds a research dinner that I would host for 10 to 15 people at the conference.
I lead SPEC-GAP, an AI safety project at BASE (Black in AI Safety and Ethics) Research. We showed that looking inside a model can reveal a hidden attack but adds almost nothing to predicting whether the model will act on it. We used probes, which are simple classifiers that read a model’s internal activity, in our case the residual stream at each of the model’s 64 layers. With them, we detected the attack even at the last agent in the chain, which never saw the attacker’s document, with an AUROC of 0.97 in the two-agent chain and 0.84 in the three-agent one. When we tried to predict which attacks that agent would carry out, the probes added little to what the test setup already told us: across the 98 runs where the attack was not resisted, they moved AUROC from 0.898 to 0.914. That prediction is what a monitor needs to know when to intervene.
The attacks we study are instructions hidden in documents, which is known as indirect prompt injection. AI systems today often split a job across several agents. In our test system, shown below, a planner assigns a task, a worker reads the documents and writes a summary, and an executor uses a tool. A hidden instruction in one document can pass from agent to agent until the executor acts on it.
Our test system. The poisoned document enters at Worker 1 and reaches the Executor either directly (2-hop) or through Worker 2 (3-hop). Each attack ends as executed, propagated (passed along without being executed), or resisted.
We ran 108 of these attacks against the Qwen3-32B model, and the system resisted only 10. Of the rest, 48 ended in a simulated unauthorized tool call (executed), and the agents passed the other 50 along without executing them (propagated). Adding Worker 2 between the first worker and the executor lowered the execution rate from 67% to 22%, though just as many instructions still made it through, about 91% at both depths. When we turned on the model’s reasoning mode, execution rose from 31% to 57%.
Links:
Paper, ‘Separating Injection Presence from Executor Compromise in Residual-Stream Probes of Multi-Agent LLM Pipelines’: https://openreview.net/pdf?id=pOJZaW2o0r
LessWrong post: https://www.lesswrong.com/posts/gLhnc4eDCGzNhkrX3/what-happens-when-a-collusion-probe-only-finds-a-thin-signal
Code: https://github.com/base-research-lab/spec-gap
I’ll present our poster on December 11th. A reviewer rated the paper a 7 (good paper, accept) and wrote that “the core message is real, but the evidence base is narrow for a stronger venue”: we tested one model, and our probes might be detecting the added text not the attack itself. In response, we are running three new experiments for the final version due November 30th:
We are repeating our tests on three more models (Qwen3.5-9B, Qwen2.5-72B, and Gemma 4 31B) to see whether the results hold at other sizes and in a different model family.
We are rewording and reformatting each hidden instruction to see whether the attack survives, and adding a harmless note as a control to check whether our probes detect the attack itself or only the extra text.
We are running the reworded attacks on the new models to see whether the two effects interact.
This is the second time outside feedback has changed the project. The first was when we showed our pilot results to researchers at FAR.AI. Two of their questions were:
“Have you implemented a behavioral sanity check to see whether the dataset [you] plan on using contains a real signal, or did [you] only show the stack works?”
“Do you plan on scaling your model to something newer that could provide more realistic data?”
The current paper is our answer: it uses a larger model and labels every run by what the agents did.
In Paris, I’ll also attend workshops mainly on agent security to see how others test security in tool-using agents. Two I want to attend are Foundations of Language Model Security: Theory, Practice and Fundamental Limits and SLM-Agents: Workshop on Small Language Models for Agentic Systems.
I also hope to meet collaborators and mentors for our group. If the dinner is funded, I’ll invite a small group of researchers and industry professionals for a longer conversation about the current state of AI safety and biosecurity work. I want to return from Paris with feedback on our results, people to work with, and methods I didn’t know about.
We’ll submit the final version of the paper with the three new experiments by November 30th and post it on arXiv. I’ll present the poster on December 11th. Within 30 days of the conference, I’ll post an update here with feedback on our results, who I met, one method from the workshops we plan to try, and at least one research decision from the trip. If the dinner is funded, the update will also say how many people came.
I would be the only one traveling. The full budget is $5,500.
Round-trip flight, San Francisco Bay Area to Paris: $1,400
Conference registration: $405
Hotel: $1,600
Meals: $400
Airport transfers and local transport: $200
Poster printing: $60
Buffer for travel costs and exchange rate: $435
Research dinner (venue, food, and drinks): $1,000
I’m a biomedical scientist moving into AI/ML, with a PhD from Ohio State and a postdoc at Stanford. My interests are AI safety and biosecurity.
My teammates on SPEC-GAP are Ifeoluwa Jayeola, Oudoum Ali Houmed, Eyiara Oladipo, Elena Ajayi, Chijioke Ubajaka, Anagha Late, and Lawrence Wagner at BASE Research: https://www.baseresearch.org/
We started with a smaller pilot, which we wrote up on LessWrong in July. The probe in that pilot barely did better than guessing. For the current paper, we moved to agents that can take actions. We built the test system in LangGraph and ran it on H200 GPUs through Modal. The study covers 27 document conditions across two delegation depths and two reasoning modes, each with a matching clean run, which comes to 216 runs and 756 agent steps with activations recorded. We tried both supervised and contrast-based probes and tested them on documents and subject areas they had not been trained on. The code is public on GitHub. The workshop reviewer described the paper as “solid, honest, well-scoped empirical work with a clear negative-ish result that is useful to the field.”
The biggest risk is that I don’t raise enough or the money comes too late, and I can’t attend in person. The dinner could also have low turnout given people’s schedules that week. I’ll send the invitations as early as I can.
On the research side, the three new experiments may not all be finished by November 30th. If one isn’t, the final version will go out without it. They may also show that our results don’t hold on other models or when the attacks are reworded. If so, we’ll report that too.
$0. We are an independent group and have paid for the project ourselves. We have applied for a BlueDot Impact Rapid Grant, and a decision is pending.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!